Voxel-based texture similarity networks reveal individual variability and correlate with biological ontologies

The human brain is organized as a complex, hierarchical network. However, the structural covariance patterns among brain regions and the underlying biological substrates of such covariance networks remain to be clarified. The present study proposed a novel individualized structural covariance network termed voxel-based texture similarity networks (vTSNs) based on 76 refined voxel-based textural features derived from structural magnetic resonance images. Validated in three independent longitudinal healthy cohorts (40, 23


Introduction
Along with mammalian evolution, the human brain, including the neocortex, allocortex, and subcortical areas, has a more complex and hierarchical organization than the brains of other mammals (Kaas, 2013;Molnar et al., 2014).Brain development is associated with changes in cytoarchitectural and synaptic organizations (Casanova and Casanova, 2019;Dori et al., 2022), which eventually shaped the brain into a complex network with small-world properties (Bullmore and Sporns, 2009) and high heritability (Miranda-Dominguez et al., 2018;Sinclair et al., 2015).A branch of brain network organization, a structural covariance network (SCN), was proposed based on structural magnetic resonance imaging (sMRI) (He et al., 2007), which measures the inter-regional covariance patterns of diverse morphological metrics, such as cortical thickness (He et al., 2008;He et al., 2007), gray matter volume (GMV) (Montembeault et al., 2012;Montembeault et al., 2016;Yao et al., 2010), cortical gyrification (Ajnakina et al., 2021;Palaniyappan et al., 2015), and surface area (Raucher-Chene et al., 2020).SCNs have significant potential for uncovering global-and local-level morphological network abnormalities in multiple neuropsychiatric disorders, including autism (Bethlehem et al., 2017), schizophrenia (Prasad et al., 2022a, b), and Alzheimer's disease (He et al., 2008;He et al., 2007).However, until recently, it was standard practice to construct SCNs based on the statistical correlation of regional morphological measures between pairs of brain regions across participants (i.e., considering subjects as the time points in correlation) (He et al., 2007;Wu et al., 2017), which is constrained by sample size and fails to capture individual variability (Cai et al., 2023).
In contrast, recent improvements in individual-level SCN (ISCN) approaches have shown potential for revealing individual covariance variability for healthy participants (Tijms et al., 2012) and further exploring covariance disruptions in patients with neuropsychiatric diseases (Han et al., 2023a;Han et al., 2023b;Liu et al., 2021;Shen et al., 2021).These ISCN approaches measure either the inter-regional correlation between known structural features, for example, morphometric similarity networks (MSN) (Li et al., 2021a;Seidlitz et al., 2018); inter-regional mutual information between probability density functions (PDF) of a structure feature, for example, the Jensen-Shannon divergence-based morphometric brain network (JSD-MBN) (Li et al., 2021b;Ruan et al., 2023b;Wang et al., 2016a); or the perturbation-based differential SCN relative to a normative SCN template, for instance, individualized differential structural networks (IDSCNs) (Han et al., 2023a;Liu et al., 2021).However, these individualized approaches were restricted either by a priori hypothesis (e.g., an IDSCN must define a group-wise SCN norm, which may be influenced by sample size, site bias, and population heterogeneity) or inadequate structural information (e.g., MSNs only use several known structural features, and JSD-MBNs only use the PDF distribution of one structural measure).
Radiomics is a powerful method for extracting hundreds of highdimensional features characterizing image spatial textures (Lambin et al., 2017), which potentially provides more plentiful biological insights beyond conventional morphometric measures, such as GMV and cortical thickness (Kim et al., 2021).These radiomic features could provide abundant covariant information between regions and thus might be considered data-driven candidates for constructing ISCNs.Two pioneer studies have reported a regional radiomics similarity network based on the multiple radiomic features of predefined regions-of-interest (ROIs) from sMRI data, which showed high reliability (Zhao et al., 2021) and revealed topological abnormalities in mild cognitive impairment patients (Zhao et al., 2022).Unlike brain tumors with clear signal boundaries, determining the anatomical boundaries between human brain subregions, especially at the individual level, is difficult.Accordingly, the ROI definition based on a common group atlas could introduce bias during the quantification of radiomic features, leading to ISCNs having reduced test-retest reliability.Thus, we speculated that constructing radiomic features at the voxel level could bridge this gap and enhance ISCN reliability.
ISCNs developed using radiomic features face significant challenges in biological interpretability.Unlike traditional structural measures with well-defined biological meanings (e.g., GMV and cortical thickness) (Montembeault et al., 2012;Montembeault et al., 2016;Yao et al., 2010), radiomic features rely solely on a data-driven approach to characterize the spatial textures of brain images, making their biological significance unclear (Lambin et al., 2017), although studies have found that radiomic measures could characterize the molecular, immunohistochemical, and pathologies of brain tumors (Park et al., 2019;Tomaszewski and Gillies, 2021) and neurodegenerative diseases (Lee et al., 2020a;Rahmim et al., 2017;Zhao et al., 2023).At the network level, researchers have attempted to investigate whether ISCNs could predict individual variability in behaviors (Li et al., 2023b;Seidlitz et al., 2018;Tijms et al., 2014), age (Li et al., 2023a;Li et al., 2023b;Ruan et al., 2023a), and sex (Li et al., 2023c), potentially strengthening the biological explanation of ISCNs.Nevertheless, it remains unclear whether and to what extent radiomic-based ISCNs can explain individual variations in behavioral and demographic biological characteristics.
Similarly, clarifying the underlying molecular biological mechanisms to enhance the understanding of the biological substrates of radiomic-based ISCN is essential.Previous neuroimagingtranscriptomics association studies have identified a significant association between transcriptional profiles and ISCN organization (Seidlitz et al., 2018) or ISCNs' disruption by disease (Liu et al., 2023;Morgan et al., 2019).Moreover, a recent study demonstrated a significant correlation between radiomic-based ISCNs and gene expression similarity networks (GESN) (Zhao et al., 2021).Although the subnetworks of SCNs have been reported to be associated with a coexpression module of a human supragranular enriched gene set (Romero-Garcia et al., 2018), it was unclear which gene set within a coexpression pathway selectively correlated with which subnetwork of radiomic-based ISCNs.Finally, the biological processes contained different scales of ontologies, including cytoarchitecture, neuroreceptors, metabolism, electrophysiology, large-scale functional, and anatomical organization (Markello et al., 2022).Elucidating radiomic-based ISCNs with these biological ontologies could significantly strengthen the understanding of their biological substrates.
In this study, we extended the radiomic-based ISCN by calculating 3D voxel-wise feature maps at the voxel level and further leveraged these texture feature maps to construct an ISCN termed voxel-based texture similarity network (vTSN).Recent advances in 3D voxel-wise texture analyses may minimize the effects of artificial ROI definitions on feature calculation (Ishaque et al., 2019;Maani et al., 2016;Ta et al., 2020).We compared radiomic-based ISCNs constructed using different feature calculation strategies, specifically the vTSN versus the regional-based texture similarity network (rTSN).Based on three longitudinal sMRI datasets, we hypothesized that the proposed vTSN would enhance test-retest reliability and improve predictions of brain transcriptional profiles.Additionally, we sought to elucidate the hidden hierarchical structures within the vTSN by examining its association patterns with various scales of biological ontologies.Finally, utilizing two large-sample sMRI datasets, we aimed to associate vTSNs with individual variations in behavior, age, and sex.

Test-retest datasets
In this study, three longitudinal sMRI datasets were recruited to estimate the ability of both types (vTSN vs. rTSN) of texture similarity networks (TSNs) to reveal individual variability, their test-retest reliability, and to associate them with biological ontologies: (1) Recruitment of the local test-retest dataset was approved by the Medical Research Ethics Committee of Tianjin Medical University General Hospital, and all subjects provided written informed consent.Relevant institutional review boards also approved the two public testretest datasets, and detailed recruitment information was provided on the website.

Cross-sectional datasets
Two cross-sectional sMRI datasets were recruited to associate vTSNs with individual variations in behavior, age, and sex.First, 424 unrelated healthy participants (188 males, age range: 22-36 years) from the Human Connectome Project (HCP) dataset (https://db.humanconnectome.org/) were enrolled to link vTSNs with individual variabilities of behaviors and sex.Since the HCP dataset contained participants of a limited age range, an additional 494 healthy participants (187 males, age range: 19-80 years) from the Southwest University Adult Lifespan Dataset (SALD) (Wei et al., 2018) were enrolled to investigate the impact of aging on vTSNs through the lifespan.

SMRI acquisition and preprocessing
The sMRI data of all healthy subjects were acquired using T1weighted 3D inversion recovery prepared gradient sequences on 3 Tesla MRI scanners.Detailed information on data acquisition is provided in Supplementary Table S1.
All sMRI data were preprocessed using a longitudinal pipeline based on CAT12 toolbox version r1364 (http://dbm.neuro.uni-jena.de/cat/)implemented in Statistical Parametric Mapping (SPM) software package version 12 (http://www.fil.ion.ucl.ac.uk/spm/) in the following way: (1) For each subject, the sMRI data from all time points were first realigned using inverse-consistent rigid-body registration, and intrasubject bias-field correction was performed.(2) Local intensity transformation (LAS) was used to correct the tissues' signal inhomogeneities.
(3) An adaptive Maximum A Posterior segmentation technique and a Partial Volume Estimation were applied to estimate the fraction of each tissue class in each voxel, including gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF).(4) A native brain mask of each subject was generated by merging only the individual GM and WM components (GM + WM > 0.95).The brain mask and the intensitynormalized brain sMRI map were used to calculate the radiomic features (see the section below).(5) An inverted nonlinear deformation map was estimated for each time point image using diffeomorphic anatomical registration through the exponentiated Lie algebra (DAR-TEL) algorithm, and then the maps from all time points were averaged.(6) Finally, individual GM and WM maps were normalized into the Montreal Neurological Institute (MNI) space using the average deformation map, and the baseline deformation map was used to unwarp the atlas in the MNI space into the subject's native space for vTSN and rTSN construction.
In addition, sMRI data from HCP and SALD datasets were preprocessed using a cross-sectional pipeline with the same parameters as the longitudinal pipeline, except for the coregistration steps.Detailed information was provided in the Supplementary Methods-sMRI preprocessing of HCP and SALD datasets.

Construction of vTSN and rTSN
All the radiomic features of sMRI were calculated in the subject's native space using Pyradiomics (van Griethuysen et al., 2017) version v2.2.0 (https://github.com/AIM-Harvard/pyradiomics)implemented in Python software version v3.6.A recent preliminary study showed that a vTSN based on only Gray Level Co-occurrence Matrix (GLCM) features could reliably reveal unique structural heterogeneity and complex dysconnectivity in patients with schizophrenia (Ding et al., 2024).This study further expanded vTSN construction by incorporating predominant types of radiomic features, including 92 candidate radiomic features across six radiomic categories: (1) first-order texture: contains first-order features; (2) GLCM: contains 23 texture features; (3) Gray Level Dependence Matrix (GLDM): contains 14 texture features; (4) Gray Level Size Zone Matrix (GLSZM): contains 16 features; (5) Gray Level Run Length Matrix (GLRLM): contains 16 features; and (6) Neighboring Gray Tone Difference Matrix (NGTDM): contains five features.Detailed information on each radiomic feature and type of radiomic feature are described in Supplementary Table S2 and Supplementary Methods: The radiomic features calculation.Shape-based features were not included because they characterized the boundary properties of a region, which was unnecessary for constructing vTSNs in a voxel-wise manner.Moreover, two popularly used atlases were introduced to define the nodes for vTSNs and rTSNs, including the Human Brainnetome Atlas (BNA, http://atlas.brainnetome.org/)and Automated Anatomical Labeling (AAL, https://www.gin.cnrs.fr/en/tools/aal/)atlas.The two atlases were unwarped into each subject's native space using a DARTEL inverted nonlinear deformation map.The fundamental difference between the vTSNs and rTSNs was their radiomic feature calculation strategies.
For vTSNs, a 3D voxel-wise radiomic algorithm was used to calculate 92 texture feature maps for each subject.Specifically, a 5 × 5 × 5 voxel cube centering on one voxel (kernel radius = 2) was extracted from the intensity-normalized sMRI map within the brain mask.Then, all values in the cube were scaled to eight gray-level intensities (binCount = 8), and 92 texture features for this voxel were derived with default settings, including distance to neighbor = 1, no normalization, no resampling, no wavelet, and no re-segmentation.The above step was repeated for each voxel, resulting in 92 texture feature maps for each subject.Then, the two brain atlases (AAL and BNA) were applied to these texture maps to extract the average values across voxels within each brain region.Since the features' scales varied dramatically from 10 − 3 to 10 2 , a z-score transformation was conducted across brain regions for each feature type to eliminate their effects on vTSN calculation.
For the rTSNs, an sMRI brain mass was first extracted for each region using the AAL or BNA atlas in native space.Next, the grayscale values of all voxels in each region mass were scaled to eight gray-level intensities, and 92 texture features were derived using default settings.Then, a zscore transformation was undertaken across brain regions for each feature type before the rTSN calculation.
Considering the possibility of high collinearity between different radiomic types (first-order, GLCM, GLDM, GLSZM, GLRLM, and NGTDM), the vTSN of each radiomic type was first constructed, and their similarities were estimated as follows: (1) For each subject, the Pearson correlation was applied to measure the covariance coefficient of the feature vectors for each texture type between each pair of brain regions, resulting in a vTSN for each texture type (Supplementary Fig. S1).
(2) The similarity between different types of vTSNs was estimated by calculating the spatial Pearson correlation between each pair of average vTSN matrix types.It was found that GLRLM-derived vTSNs were similar to GLDM-(Rs > 0.94) and GLCM-derived ones (Rs > 0.87) across different atlases and datasets.Similar patterns were identified for rTSNs (Supplementary Fig. S2).We speculated that GLRLM features contribute little to a vTSN.Thus, GLRLM features were excluded from the construction of vTSN and rTSN to avoid redundancy.Finally, radiomic features were incorporated into calculating vTSNs or rTSNs for each subject using Pearson correlation.
L. Lin et al.

Construction of the network topology of vTSNs and rTSNs
The graph approach provides a compelling means of characterizing brain network topological architectures.To investigate the ability of vTSNs and rTSNs to detect individual topological heterogeneity, the vTSN and rTSN network metrics were computed using the Graph-Theoretical Network Analysis (GRETNA) toolkit (Wang et al., 2015), including both global (shortest path length [Lp], clustering coefficient [Cp], normalized clustering coefficient [γ], normalized shortest path length [λ], small-world properties [σ]), and nodal measures (nodal clustering coefficient, ci; nodal local efficiency, locEi; nodal global efficiency, gEi; nodal degree, ki; and nodal betweenness, bi) (Rubinov and Sporns, 2010;Wang et al., 2011).Regarding the different biological substrates of the positive and negative SCNs indicated by a previous study (Gong et al., 2012), the original or absolute values of the TSN matrix in graph analyses were not used to establish topological properties.Instead, negative TSN values were eliminated (setting them to zero), and positive connections were solely employed to construct the graph under various sparsity thresholds, aligning more closely with the tractography-based anatomical network (Gong et al., 2012).Notably, many prior topological analyses of SCNs have adopted this approach (Hosseini et al., 2016;Mareckova et al., 2022;Wang et al., 2016b;Zhang et al., 2019).A wide range of sparsity from 0.05 to 0.5, with a step of 0.05, was applied to calculate global and nodal network measures at each sparsity level, and the area under the curve (AUC) for each network metric was computed, which provided a sparsity-independent summary scalar for the topological characterization of brain networks.Moreover, during small-world metrics (γ, λ, σ) calculation, 1000 random networks were generated to obtain the denominator of each metric.

Quantifying intra-and inter-subject variability of vTSNs and rTSNs
Inter-subject and intra-subject variability in the three longitudinal test-retest datasets were estimated and compared.The area-level intrasubject variability of ROI i and subject s (Wsi) was quantified as equation [1]: where corr is the function of Pearson's correlation, and m; n = 1, 2 …, M (m ∕ = n), with M representing the number of sessions.
By averaging the 246 area-level variabilities across the whole brain, we obtained participant-level intra-subject (Ws) and inter-subject variability (Vs): (3) We also introduced the connectivity variability ratio (VR) to quantify the ability of vTSN/rTSN to detect inter-subject heterogeneity after controlling for intra-subject variability: The calculation of intra-and inter-subject variability for network metrics was similar to that of the variability derived from the vTSN matrix, specifically, by replacing the TSN i (s, t) of equation [1] and [2] with the network metric vector (NMV) (s,t), where NMV(s, t) represents the five z-score normalized nodal network metrics across brain regions (ci, Lpi, locEi, gEi, ki, bi).

Estimating the test-retest reliability of vTSNs and rTSNs
Test-retest reliability is a statistical strategy commonly used to assess the consistency or stability of a measure over time.Reliable test-retest measurements were crucial for drawing convincing conclusions, and only sufficiently reliable measures could serve as candidate imaging biomarkers for brain science research.Consequently, the intra-class correlation coefficient (ICC) (Shrout and Fleiss, 1979) was also calculated to quantify the test-retest reliability of vTSNs and rTSNs and their topological properties.The ICC was defined as follows: In this equation, MS P was the between-participants mean square of each vTSN and rTSN measure, MS w was the within-subject mean square, and k represented the number of repeated measurements.

Gene expression data preprocessing
The Allen Human Brain Atlas dataset (AHBA, http://human.brain-map.org/)contained microarray expression data collected from six healthy donors (age = 42.50 ± 13.38 years; male/female = 5/1) with 3702 spatially distinct samples.We mapped AHBA samples to parcels of the BNA brain atlas using the abagen toolbox (https://www.github.com/netneurolab/abagen).The preprocessing of gene expression data was summarized as follows: (1) updating probe-to-gene annotations using the information provided by Arnatkeviciute et al. (Arnatkeviciute et al., 2019); (2) probe intensity filtering (probes with intensity less than the background in ≥ 50 % of samples were discarded); (3) probe selection (selecting representative probe among the probes indexing the same gene according to differential stability); (4) matching tissue samples to regions defined in the BNA brain atlas; (5) sample expression values separately normalized for each sample and donor across genes; (6) gene expression values normalized separately for each gene for each donor across all expression samples; (7) aggregating samples within regions based on matches made in step 5; and (8) gene set filtering based on differential stability, with only the 50 % of genes with the highest differential stability score (N genes = 7817, DS > 0.3476) being used for subsequent analysis (Hawrylycz et al., 2015).Lastly, a gene expression matrix with 246 regions × 7817 genes was obtained for further analyses.

Identifying coexpression gene modules
A weighted gene coexpression network analysis (WGCNA) was applied to identify coexpression modules whose genes showed similar coexpression profiles (Langfelder and Horvath, 2008) in the following way: (1) a signed coexpression matrix was constructed between all pairs of genes across brain regions using Pearson's correlation, resulting in a symmetric matrix (7817 × 7817); (2) A soft-thresholding power of was chosen to construct the adjacency matrix based on a scale-free topology criterion (fit index > 0.8) to highlight the differences between strong and weak correlations between gene pairs; (3) the transformed matrix was converted to a Topological Overlap Matrix (TOM) to minimize the effects of noise and spurious associations; (4) a hierarchical clustering was applied to the TOM to identify gene modules.Thirteen meaningful gene coexpression modules were identified from clustering, including the following: black (n = 233 genes), blue (n = 441), brown (n = 298), cyan (n = 163), gray60 (n = 59), green (n = 560), greenyellow (n = 125), lightcyan (n = 81), purple (n = 238), red (n = 234), tan (n = 115), turquoise (n = 701), and yellow (n = 286), leaving out the gray module composed of all unassigned genes (n = 4283).Finally, the module eigengene (ME), a representative of the gene expression profile of each coexpression module, was derived using the first principal component.

Test-retest reliability comparison between the vTSN and rTSN
First, to explore whether the participant's vTSNs and rTSNs could discriminate them from others, a non-parametric Wilcoxon signed-rank test was performed to compare the inter-subject and intra-subject variability of the vTSNs and rTSNs at both the area level (P < 0.05/246 regions, Bonferroni correction) and the participant level (P < 0.05/3 datasets/2 atlases, Bonferroni correction).To compare rTSNs and vTSNs in terms of their ability to detect individual heterogeneity as well as the test-retest reliability, we used a non-parametric Wilcoxon signed-rank test to compare differences in VR and ICC between vTSNs and rTSNs, respectively (P < 0.05/3 datasets/2 atlases/2 metrics, Bonferroni correction).

Associations between vTSNs and rTSNs and genome-wide GESN
To address the underlying association between gene expression and vTSNs and rTSNs, the vTSNs and rTSNs were initially averaged across subject sessions and datasets, resulting in mvTSNs and mrTSNs.A gene expression similarity network (GESN) was constructed using Pearson's correlation across all 7817 gene expression profiles between each pair of brain regions.Then, a partial correlation coefficient was used to estimate the couplings of mrTSN-GESN and mvTSN-GESN, controlling for inter-regional Euclidean distance to remove spatial autocorrelation.The cocor toolbox (Diedenhofen and Musch, 2015) was used to compare the correlation coefficients between mrTSN-GESN and mvTSN-GESN in consideration of the coefficients' dependency (Meng. et al., 1992) (P < 0.05).We also validated the couplings in each dataset (P < 0.05/3 datasets, Bonferroni correction).To investigate the extent to which each type of the five radiomic types contributed to the vTSN-transcription associations, we also estimated the couplings of the vTSN of each radiomic type (first-order, GLCM, GLDM, GLSZM, GLRLM, and NGTDM) and the GESN.

Associations between nodal vTSNs and coexpression eigengenes
To explore whether the hierarchical organization of vTSNs was selectively regulated by specific coexpression gene modules, we further tested whether the vTSN profile of each brain region was associated with the eigengene of each coexpression gene module.For each gene module and brain region, we calculated partial correlation coefficients between the module eigengene and the area-level vTSN, controlling for the Euclidean distance of this brain region with the remaining 245 covarying regions.Thus, a 246 × 13 vTSN-eigengene correlation matrix was generated.
A hierarchical agglomerative clustering method was applied to the vTSN-eigengene correlation matrix to separate brain regions into independent subnetworks, with regions within a subnetwork that had similar vTSN-coexpression association patterns that differed from others outside the subnetwork.Euclidean distance was used to estimate clustering distance, and the "farthest distance" method was used to join most similar clusters.NbClust (Charrad. et al., 2014) was used to determine the optimal cluster numbers in which the brain regions should be grouped.This tool provides 30 types of clustering number evaluation rules, such as the Calinski-Harabasz score and the silhouette coefficient.All 30 rules were tested, and the optimal cluster numbers were determined to be three according to the majority rule (12 criteria support three clusters as the best number of clusters).Therefore, 246 brain regions were split into three large subnetworks according to vTSN coexpression associations.To finely sub-divide these three subnetworks, clustering analysis was conducted on each of them, using the same algorithms.The clustered vTSN-coexpression subnetworks are publicly available at the GitHub repository (https://github.com/BrainWanderLab/vTSN_Subnets.git).

Coexpression gene module enrichment analysis
To identify the potential biological functions of each gene module, GO term enrichment analysis was performed with the annotated genes as the background on web server g: Profiler (https://biit.cs.ut.ee/ gprofiler).Cell-type enrichment analysis was also conducted using the specificity index probability (pSI) package in the R program (Dougherty et al., 2010) to explore which cell types were specifically expressed in these gene modules.

Associations between nodal vTSNs and biological ontology
To correlate vTSNs with different scales of biological ontology, including neurotransmitters, metabolism, electrophysiology, and the large-scale organization of brain function and structure, a series of highresolution biological ontology maps derived from the neuromaps software toolbox were used (Markello et al., 2022).Of these neuromaps, two related to evolutionary and developmental expansion (Hill et al., 2010) were excluded because they covered only half of the hemisphere.Additionally, brain maps related to cognition (1) and gene expression (1) were also excluded, as these topics were comprehensively explored in the analyses.In the remaining 46 biological maps, the same receptor and transporter maps (e.g., 5-HT1B, D2, mGluR5, and VAChT) were averaged following the preprocessing pipeline recommended by Hansen et al. (2022).A total of 40 biological ontology maps were included in the subsequent association analyzes comprising 19 neurotransmitters, electrophysiology, 4 metabolite, and 6 large-scale structural, and functional organization maps.Details are provided in Supplementary Table S3.
Following the main body of this study, BNA was used to extract the regional values of each biological ontology map.Since most of these biological brain maps only contain the cortex, we only chose the cortical parcellations to link vTSNs with biological ontology across regions.Finally, a 210 × 40 vTSN-biological ontologies correlation matrix was obtained using partial correlation, controlling for inter-regional Euclidean distance (P < 0.05/ 210 regions/ 40 biological ontologies, Bonferroni correction).
To clarify the biological framework underlying the correlation patterns between the vTSNs and biological ontologies, a hierarchical agglomerative clustering approach was employed on the vTSNbiological ontology correlation matrix, and the optimal number of clusters were evaluated using the NbClust package in the R program.

The correspondence between vTSNs and cytoarchitecture
To further test the biological validity of vTSN subnets, the correspondence between the vTSN and cytoarchitecture classes proposed by Von Economo and Koskinas were explored (von Economo and Koskinas, 1925).They categorized the brain cortex into six cytoarchitectonic classes: agranular, frontal, parietal, polar, granulous, and transition cortices.This cytoarchitectural atlas was digitized by Pijnenburg et al. (Pijnenburg et al., 2021).BNA parcellations were first ascribed to the cortical class that had the most overlapping voxels out of the six cortical classes.Subsequently, a series of sparsely connected graphs were created using the average vTSN from three test-retest datasets (sparsity ranging from 0.5 % to 5 % in 0.5 % increments).The densities of surviving intra-class and inter-class connections were calculated by dividing the number of surviving intra-class connections by the total intra-class connections, and surviving inter-class connections by the total inter-class connections, respectively.Finally, the density ratio between intra-and inter-class connections was obtained.

Associations between vTSNs and individual variability in human behavior
A multi-step screening strategy was used to preprocess behavioral measures of the HCP dataset (Li et al., 2023b).Sixty items were retained and classified into six behavioral domains: alertness, cognition, emotion, motor, personality, and sensory.Furthermore, 17 participants from this study were excluded because they missed either one of the 60 behavioral items, resulting in 407 participants for vTSN-behavior association analysis.To eliminate the outlier effects, behavior values that deviated by 3.29 times the standard deviation (beyond the 99.9 % confidence interval) were replaced by this extreme value.Finally, each behavior item underwent z transformation, and all items within each behavioral domain were averaged, resulting in six integrated behavioral metrics.
A multivariate variance component model (MVCM) was introduced to investigate the potential of vTSNs for explaining inter-individual behavioral variability (Ge et al., 2016).Specifically, the total variance contribution of global-and subnet-level vTSNs on integrated behavior metrics was explored.In the MVCM model, each integrated behavior metric was defined as the dependent variable, and the across-subject similarity matrix (ASM) variable was defined as the independent variable, representing the similarity of vTSNs between each pair of subjects.
Based on the optimal clusters derived from vTSN-coexpression association patterns, the whole brain vTSN was subdivided into the within-subnet and between-subnet vTSNs.ASMs for each subnet-level vTSN were also calculated, and their contribution to behavioral individual heterogeneity was explored.A permutation test was used to estimate the significance of the explained variance of vTSNs on each behavior domain (1000 permutations) and the false discovery rate (FDR) correction was applied to correct multiple comparisons (q < 0.05).

Effects of age and sex on vTSNs
The Spearman correlation was used to test the associations between age and the vTSNs based on 494 healthy adults from the SALD dataset, controlling for the factor of gender (P < 0.05/30,135, Bonferroni correction).Moreover, a two-sample t-test was used to test the sex differences (males vs. females) in vTSNs after regressing out the age effects (P < 0.05/30,135, Bonferroni correction).

Associations between the vTSN/rTSN and the genome-wide GESN
The underlying association between the GESN based on all qualified genes and the vTSN/rTSN (Fig. 5a) was further explored.The partial correlation coefficient revealed a significant positive association between the mean vTSN and the GESN of all qualified genes (r = 0.600, P < 0.001) and between the mean rTSN and the GESN (r = 0.433, P < 0.001) (Fig. 5b), which was reproducible in each test-retest dataset (Fig. 5c), indicating that region pairs with high morphometric similarity also tended to have high transcriptional similarity.Moreover, the correlation coefficient comparison results showed that vTSN-GESN association was greater than that of rTSN-GESN (z = 39.784,P < 0.001) (Fig. 5b), which was also validated in each dataset (Local: z = 40.276,P < 0.001; Duke: z = 39.316,P < 0.001; BNU: z = 38.383,P < 0.001) (Fig. 5c), suggesting that vTSNs performed better in predicting the transcriptional similarity between regions.Furthermore, the vTSN of the GLCM (r = 0.615) showed the strongest association with the GESN, while the vTSN of NGTDM (r = 0.482) displayed the weakest correlation with the GESN (Fig. 5d), suggesting that GLCM possesses the most robust predictive power for the brain transcriptome.

Associations between nodal vTSN and coexpression eigengenes of gene modules
WGCNA identified 13 gene coexpression modules (Fig. 6a, b).The association patterns between the nodal vTSN profile and the coexpression eigengenes of each gene module were diverse in various brain regions (Fig. 6c).The hierarchical clustering method identified three optimal large subnetworks based on the majority rule (12/30 rules support three clusters) according to vTSN-coexpression association patterns (Supplementary Fig. S6a).Subnet-1 mostly contained brain regions within the subcortex and limbic cortex, such as the bilateral thalami, basal ganglia nuclei, insular, hippocampi, and cingulated cortices, with a vTSN profile mostly positively associated with the eigengene of brown, lightcyan, red, black, blue, green, tan, and yellow gene modules and negatively associated with the greenyellow, gray60, purple, and turquoise gene modules.Subnet-2 mostly contained the brain regions in the dorsal neocortex, including all bilateral primary sensory areas (somatosensory, visual, and auditory), dorsal visual and parietal cortices, and lateral prefrontal cortices, and presented the opposite vTSN-coexpression association than Subnet-1 did.Subnet-3  mainly contained the brain regions in the ventral neocortex, including the bilateral temporal, ventral, and medial prefrontal cortices.The vTSN-coexpression similarity pattern of Subnet-3 was predominantly fell between that of Subnet-1 and Subnet-2 and was more closely resembling dorsal neocortex (Subnet-2).Except for the yellow, blue, and cyan modules, Subnet-3 showed a closer association than Subnet-2 (Fig. 6c, d).The three subnetworks could be further split into 10 refined ones (three for Subnet-1, four for Subnet-2, and three for Subnet-3) based on the majority rule (Figs.6e, S6b-d), whose vTSN profile demonstrated relatively unique association patterns with the coexpression profiles of the 13 gene modules.

Gene module enrichment analysis
GO enrichment analysis was conducted for each gene module.GO ontology terms found that the brown, lightcyan, red, black, blue, and green gene modules were enriched in developmental process regulation, immune system process, and immune system process pathways, and the greenyellow, gray60, tan, yellow, purple, cyan, and turquoise gene modules were enriched in pathways related to synaptic signaling, cell communication, regulation of the biological process, the protein metabolic process, and potassium ion transmembrane transport (P <0.05, FDR correction) (Fig. 7a-c; Supplementary Tables S6).Moreover, celltype enrichment analyses showed that the brown, light cyan, red, blue, and green gene modules were mainly enriched for glial cells, and the green-yellow, tan, yellow, purple, and cyan gene modules were specifically enriched for neurons (Fig. 7d, Supplementary Tables S7).

Associations between vTSNs and biological ontology and cytoarchitecture
The Partial correlation demonstrated that except for six biological ontologies (three cortical areal scaling, A4B2, CB1, VAChT, and evolutionary cortical expansion), the remaining 33 biological ontologies had significant correlations with vTSNs in either one of the 210 brain regions (P < 0.05/210 regions/40 biological ontologies, Bonferroni correction).Hierarchical agglomerative clustering identified two optimal subsets of brain biological ontologies based on the majority rule (9/30 rules support two clusters).Biological ontology type-1 predominantly included electrophysiological activity and large-scale brain organization, whereas biological ontology type-2 primarily comprised neurotransmitters, metabolism, and myelination.The vTSN profiles of different subnetworks exhibited distinct correlation patterns with these biological Fig. 7. Enrichment of genes in 13 gene modules.Gene ontology (GO) enrichment for (a) molecular functions, (b) cellular components, and (c) biological processes (top three ontology terms shown for each module for molecular functions, cellular components, and biological processes).(d) Cell-type enrichment of genes in 13 modules.All obtained results could pass q < 0.05 (BH-FDR correction).ontology types.For instance, the profiles of biological ontology type-1 were mostly positively correlated with the subcortex-limbic system vTSN and negatively correlated with the dorsal neocortex vTSN.Conversely, a roughly opposite pattern was observed for biological ontology type 2. Finally, the vTSN-biological association patterns for the ventral neocortex lay between the subcortex-limbic system and the dorsal neocortex (Fig. 8a).In combination with the associations between vTSN subnets and coexpression/biological ontologies, these findings indicated that the vTSN subnets had unique biological meanings, with hierarchical organization from the subcortex-limbic system to the  S6 provides information on biological ontology.ventral neocortex and dorsal neocortex.
Regarding the correspondence between vTSNs and cytoarchitecture classes, it was determined that all intra/inter-class density ratios were higher than two (from 2.17 to 3.49) under different network sparsity thresholds (Fig. 8b), suggesting that the strongest vTSN connections link brain regions of the same cytoarchitectural class.

Effects of age and sex on vTSNs
As shown in Fig. 10a, significant positive (5163 connections, involving 17.13 % of vTSN connections) and negative (5354 connections, involving 17.77 % of vTSN connections) Spearman correlations between age and the vTSN (P < 0.05/ 30,135, Bonferroni correction) were found.At the subnet level, the vTSN between the dorsal and subcortex-limbic subnets had the highest proportion of positive correlations with age (2890 connections, involving 43.46 % of the vTSN in the subnet), and the vTSN within the subcortex-limbic subnet had the highest proportion of negative correlations with age (1110 connections, involving 45.96 % of the vTSN).
A two-sample t-test demonstrated that the vTSN strength of connections (2.16 %) was greater in males than females and that connections (1.66 %) were greater in females.At the subnet level, a higher vTSN in males primarily involved connections between the ventral and subcortex-limbic subnets (4.23 %), and a higher vTSN in females mostly involved connections within the subcortex-limbic subnets (11.35 %) (Fig. 10b).

Discussion
The present study proposed a novel individualized SCN measure termed voxel-based TSN (vTSN) based on 76 refined texture feature maps.Consistent with our initial hypothesis, we validated that the proposed voxel-based vTSN not only exhibited a higher ratio of intersubject to intra-subject variability compared to the region-based TSN (i.e., rTSN), but also exhibited a higher intra-class coefficient in the testretest reliability of both connectivity strength and network topology across different datasets and brain atlases.Moreover, the vTSN also demonstrated more associations with the GESN than with the rTSN.Furthermore, we identified three vTSN subnetworks based on vTSNcoexpression association patterns with unique biological hierarchical organizations from the subcortex-limbic system to the ventral neocortex and to the dorsal neocortex.Subject-level association analyses revealed that vTSNs could explain a significant proportion of inter-subject behavioral variability in cognition, motor, and emotion.Finally, we demonstrated that vTSNs could sensitively characterize differences in brain structural networks shaped by sex and accurately depict the developmental trajectory of structural networks throughout the one's lifespan.In summary, our proposed vTSN is robust in uncovering individual variability and neurobiological brain processes, which can serve as a biologically plausible measure for linking biological processes and human behavior.
Traditional SCN approaches were constructed by taking subjects as time points at the group level (He et al., 2007;Wu et al., 2017), which had common limitations, including the need for adequate subjects to form a stable "time series" and ignorance of individual heterogeneity.Thus, individualized SCNs are expected to resolve these issues (Li et al., 2017;Seidlitz et al., 2018).Our study extended the ROI-based radiomic similarity network (Zhao et al., 2021) to a voxel-wise feature manner that calculates the inter-regional covariance using 76 refined texture maps, including first-order statistics, GLCM, GLDM, GLSZM, and NGTDM feature maps for each subject.Radiomics is a powerful, robust method for extracting a significant number of statistical and texture features from a mass (such as tumors and brain lesions) in medical imaging analysis (Chen et al., 2017;Ishaque et al., 2018;Luk et al., 2018;Maani et al., 2016).This method may provide additional latent information on the structural organization of a brain region relative to traditional structural measures, such as volume, cortical thickness, and surface area.Therefore, radiomics features could be considered potential measures for reconstructing between-region structural covariance at the subject level.Our vTSN and two recent rTSN studies (Zhao et al., 2021;Zhao et al., 2022) prove that individualized SCNs based on radiomic features have high test-retest reliability in characterizing individual heterogeneity in brain structural connectivity and could represent complex brain network topological properties (such as small worldness).
We found that the test-retest reliability of vTSNs was relatively higher than the rTSNs in both variability ratio and intra-class coefficient for either the original connectivity matrix or the network topology, which was replicated by three independent datasets and two different L. Lin et al. nodal-defined atlases.The potential explanations for the strengthened performance may include the following: (1) the vTSN extracted more detailed information from brain images because of its voxel-by-voxel analysis, meaning that it can better characterize individual heterogeneity; (2) the vTSN used a 3D voxel-wise texture algorithm rather than an ROI-based strategy to calculate radiomic features, which may minimize systematic errors caused by imperfect boundary-definition by group-wise brain atlases, thus increasing test-retest reliability (Ishaque et al., 2019;Ta et al., 2020).Additionally, we found that the vTSN at the connectivity level was generally reproducible across scans, datasets, and atlases.However, the reproducibility of vTSNs in terms of graph measures was lower than that observed at the connectivity level (particularly for locEi and bi) and was also lower than that reported in previous studies on ISCNs (Li et al., 2021b;Zhao et al., 2021), which should be used with caution.Further studies should directly compare the test-retest reliability between vTSNs and other types of ISCN, as well as between vTSNs and tractography-based anatomical networks and functional networks under the same benchmark (e.g., identical datasets).
One of the significant challenges concerning the proposed vTSN is its biological interpretability.Unlike other SCNs based on traditional hypothesis-based measures (such as GMV and cortical thickness) (Montembeault et al., 2012;Montembeault et al., 2016;Yao et al., 2010), this vTSN is constructed solely used data-driven radiomic features, with vague biological meanings (Lambin et al., 2017).To better understand the biological substrates of our proposed vTSN, we first attempted to associate it with gene expression profiles and cytoarchitectural organizations.Brain organization exhibited widespread conservation in gene expression across individuals and species (Hawrylycz et al., 2015).Moreover, the co-transcription pattern was similar in layered architecture across the entire brain, indicating that gene expression profiles strongly influence cortical development.Previous SCN studies on the group-level and individual-level (Romero-Garcia et al., 2018;Seidlitz et al., 2018) demonstrated that greater structural covariance between brain regions was associated with greater similarity in inter-regional gene coexpression.For example, it has been shown that the spatial patterns of gene expression are highly associated with the organization of MSNs (Seidlitz et al., 2018) and MSN abnormality in schizophrenia (Morgan et al., 2019).A recent study also established that the ROI-wise radiomic similarity network strongly correlated with the genome-wide mRNA expression similarity network (Zhao et al., 2021).As expected, we found that vTSNs could significantly predict the GESN (r ~ 0.6), indicating a close association between vTSNs and coexpression across brain regions.We further revealed that the predictive power of vTSNs for the GESN varies among different texture types, and the GLCM possesses the most robust predictive power for brain transcriptome.In addition, we confirmed that the vTSN demonstrated more associations with the GESN than the rTSN did.The strengthened vTSN-transcription coupling further validated the robustness of our proposed vTSN in revealing biological heterogeneity across brain regions.On the correspondence between TSN and cytoarchitecture classes (Pijnenburg et al., 2021;von Economo and Koskinas, 1925), we found that stronger vTSN connections tend to connect brain regions belonging to the same cytoarchitectural class, indicating a close relationship between vTSNs and cytoarchitectural organization that is also found in MSNs (Seidlitz et al., 2018).
A recent study of the transcriptional relevance of SCNs at the geneset level determined that SCN patterns of certain prefrontal subregions were selectively regulated by the expression similarity network of dopamine-related gene sets (Liu et al., 2023).Furthermore, the SCNs of brain regions had particularly high levels of coexpression of a human supragranular enriched gene set, which are important for large-scale, long-distance cortico-cortical connectivity (Romero-Garcia et al., 2018).In this study, we extended these works by associating the vTSN with all differential transcriptional gene sets derived from the data-driven WGCNA method.Our findings demonstrated that the vTSN of different brain regions have unique association patterns with specific coexpression modules.We also identified three major hierarchically organized subnets (the subcortex-limbic system, ventral neocortex, and dorsal neocortex) whose vTSNs were differentially associated with the 13 coexpression gene modules.For example, the vTSN of the subcortex-limbic system demonstrated positive associations with the transcription level of gene sets enriched in developmental and immune pathways of astrocyte and microglia, while negative associations with the transcription of gene sets enriched in synaptic signaling, cell communication, and potassium ion transmembrane transport pathways of neurons.However, a nearly opposite vTSN-coexpression association was found for the dorsal neocortex subnet unlike in the subcortex-limbic subnet, and the ventral neocortex network manifested a transition vTSN-transcription pattern between the subcortex-limbic system and the dorsal neocortex.We further associated these three vTSN subnets with various biological processes, such as neuroreceptors, metabolism, electrophysiology, and the large-scale organization of brain function and structure (Markello et al., 2022;Pijnenburg et al., 2021), and found that 33/40 biological ontologies had significant correlations with vTSNs in either one of the 210 brain regions, which could be further clustered into two biological categories and had hierarchical association patterns with the three vTSN subnets similar to that of coexpression gene modules.In combination with the associations between vTSN subnets and coexpression and biological ontologies, these findings indicated that vTSN subnets have unique biological meanings and follow a hierarchical organization from the subcortex-limbic -> ventral -> dorsal neocortex, although the gradient was different from that identified by functional connectome (Knodt et al., 2023;Margulies et al., 2016).To our knowledge, this study is the first to recognize SCN subnets based on SCN-coexpression similarity information, which may be considered a unique atlas for neuroscience and neuropsychiatric studies.
To better understand the biological meanings of vTSN, we sought to further elucidate the association between vTSNs and individual variability in behavioral and demographic traits.Although many studies have indicated that brain structural and functional networks may be considered the endophenotypes for behavioral traits (Lee et al., 2020b;Li et al., 2019;Liegeois et al., 2019;Lin et al., 2020;Popp et al., 2024;Siegel et al., 2016), only a few have focused on the links between ISCN and behavior (Li et al., 2023b;Seidlitz et al., 2018;Tijms et al., 2014) .We demonstrated for the first time that vTSNs can selectively account for a significant portion of the inter-subject behavioral variability in cognition, motor, and emotion functions but not in alertness, personality, and sensory perception, which was in accordance with a recent study based on JSD-MBN (Li et al., 2023b).Thus, these findings demonstrated that vTSNs can characterize certain dimensions of human behavior and have the potential to resolve connectivity disruption in behavioral disorders, such as schizophrenia (Ding et al., 2024), depression, and dementia (Zhao et al., 2022).
Regarding the associations between vTSNs and demographic characteristics, we also observed a significant association between vTSNs and age in healthy adults from 19 to 80 years old, especially in the subcortex-limbic system (positive associations) and between the subcortex-limbic system and the dorsal neocortex (negative associations), which was consistent with the results of the resting-state networks (Betzel et al., 2014).Therefore, vTSNs could be a new metric for characterizing the development trajectory of the brain's structural network throughout one's lifespan.We also observed widespread differences in vTSNs between males and females, with higher connectivity between the subcortex-limbic system and the ventral neocortex for males but higher connectivity within the subcortex-limbic system for females.A previous study reported gender differences in SCNs and Functional Connectivity Networks (Gong et al., 2011).Our findings indicate that vTSNs can also sensitively characterize differences in brain structural networks caused by sex.However, the factors contributing to gender differences in vTSNs, as well as the associations between gender differences and behavioral differences in vTSNs, still need to be clarified.
Some limitations of this study should be addressed.First, we found that the reproducibility of the graph measures of vTSNs was relatively low, indicating that the measures of the current version should be used cautiously and improved with more advanced technology.Second, we observed a close association between vTSNs and various scales of biological processes.However, these associations were established at the group level.It is crucial to elucidate the genetic and environmental factors and their direct molecular pathways in forming and shaping vTSNs based on multi-omic data at the subject level.Third, the clinical utility of vTSNs in diagnosing and guiding treatment for neuropsychiatric diseases warrants further investigation.

Declaration of competing interest
The authors declare no conflict of interest.

Fig. 1 .
Fig. 1.Differences between intra-and inter-subject variability for local test-retest dataset.Inter-subject variability (Vs, pink) is significantly higher than intrasubject variability (Ws, blue) for (a) the vTSN based on BNA, (b) the vTSN based on AAL, (d) the rTSN based on BNA and (e) the rTSN based on AAL at participant level (P<0.05/3datasets /2 atlases, Bonferroni correction).Area-level Vs are generally higher than Ws for both the vTSN (c) and the rTSN (f) (P < 0.05/246 regions, Bonferroni correction).

Fig. 2 .
Fig. 2. Differences in connectivity VR between vTSNs and rTSNs for three independent test-retest datasets.The VR of the vTSN (pink) is significantly higher than that of the rTSN (blue) at the individual level for the Local (a), Duke (c), and BNU (e) datasets (P < 0.05/3 datasets/2 atlases/2 metrics, Bonferroni correction).VR comparisons between vTSNs and rTSNs at the area level are also shown for the Local (b), Duke (d), and BNU (f) datasets (P < 0.05/246 regions, Bonferroni correction).The number on the horizontal line represents the ratio of the median connectivity VR; *: P < 0.05/3 datasets/2 atlases /2 metrics, Bonferroni correction.VR = Variability Ratio.

Fig. 3 .
Fig. 3. Differences in connectivity ICC between vTSNs and rTSNs from three independent test-retest datasets.(a) Connectivity ICC distributions for vTSNs (orange and pink) and rTSNs (light blue and blue) for the Local (left), Duke (middle), and BNU (right) datasets.(b) The ICC of the vTSNs and the rTSNs at the area level (P < 0.05/246 regions, Bonferroni correction).The number on the horizontal line represents the ratio of the median connectivity ICC value; *: P < 0.05/3 datasets/2 atlases /2 metrics, Bonferroni correction.

Fig. 4 .
Fig. 4. Intra-and inter-subject variabilities, variability ratio, and ICC of nodal network metrics.(a) Inter-subject network variability (NV, pink) was significantly higher than the intra-subject network variability (NWs, blue) in the three test-retest datasets.(b) Comparison of NV ratio between vTSNs and rTSNs.(c) Comparison of nodal network metrics ICC values between vTSNs and rTSNs.Abbreviations: ci = nodal clustering coefficient; locEi = nodal local efficiency; gEi = nodal global efficiency; ki = nodal degree; bi = nodal betweenness.The number on the horizontal line represents the ratio of the median NV ratio (b) or the ratio of median network metrics ICC values (c); *: P < 0.05/3 datasets /2 atlases /2 metrics, Bonferroni correction.

Fig. 5 .
Fig. 5. Correlation between the vTSN and rTSN with the GESN.(a) Heatmaps for the mean vTSN, mean rTSN, and the GESN.(b) Scatterplots for the correlation coefficient between the mean vTSN (across datasets) and the GESN (left), and between the mean rTSN (across datasets) and the GESN (right).(c) Scatterplots for correlation coefficients between the vTSN and the GESN and between the rTSN and the GESN for each test-retest dataset, respectively.(d) Scatterplots for correlation coefficients between each type of vTSN and the GESN (P < 0.05, Bonferroni correction).Abbreviations: GESN = gene expression similarity networks; rTSN = regional-based texture similarity networks; vTSN = voxel-based texture similarity networks; GLCM = Gray Level Co-occurrence Matrix; GLDM = Gray Level Dependence Matrix; GLSZM = Gray Level Run Length Matrix; NGTDM = Neighboring Gray Tone Difference Matrix.

Fig. 6 .
Fig. 6.WGCNA analysis of expression data and clustering results of vTSN-coexpression association patterns.(a) The scale-free topology model fit index corresponded to different soft thresholds.When the soft threshold was 19, the scale-free topology fit index was nearly 0.8.(b) Gene dendrogram was obtained using hierarchical clustering based on the topological coexpression of 7667 genes.The color row below the dendrogram shows the module assignment of the genes.(c) The heatmap represents the vTSN-eigengene correlation matrix.The hierarchical clustering method identified three optimal large subnetworks according to vTSNcoexpression association patterns.The histogram underneath the heatmap shows the average correlation coefficient of the corresponding brain regions for each brain subnetwork.(d) The brain regions contained in each of the three subnetworks.(e) The three subnetworks were further split into 10 refined ones.

Fig. 8 .
Fig. 8. Association between the vTSN and different scales of cortical biological processes and cytoarchitecture.(a) Heatmap represents the vTSN biological ontology correlation matrix (t value).The hierarchical clustering method identified two optimal biological ontology types according to the vTSN-biological ontology association patterns.The bar charts underneath the heatmap demonstrate the average correlation coefficient of corresponding brain regions for each brain subnetwork.(b) The left panel represents the intra/inter-class density ratio under a series of sparsity, and the right panel shows the survival connections with a sparsity of 0.5 %.Nodes and intra-class connections are colored according to the cytoarchitecture class, while inter-class connections are drawn in gray.Supplementary TableS6provides information on biological ontology.

Fig. 10 .
Fig. 10.The effect of age and sex on vTSNs.(a) Connections with significant aging-vTSN correlation, (b) connections with significant sex differences in vTSNs, categorized by subnets (P < 0.05/30,135 connections, Bonferroni correction).The bar charts of the right panels represent the proportion of significant connections.
a local testretest dataset collected from Tianjin Medical University General Hospital including 40 health subjects (age: 22.75 ± 2.36 years, male/ female: 19/21) with each subject undergoing two sMRI scans one day apart; (2) a public test-retest dataset containing 23 health subjects (age: 23.48 ± 3.23 years, male/female: 15/8) from Duke University (Duke, http://duke.edu/~morey005/ScanRescanData/)with four sMRI scans, two held on day one and two one week later; (3) and a public dataset from Beijing Normal University (BNU, http://fcon_1000.projects.nitrc.org/indi/CoRR/html/bnu_2.html) with 61 healthy subjects (age: 21.36 ± 0.86 years, male/female: 46/15), including two MRI scans for each subject with an interval of 103~189 days, among which one subject was excluded because of poor image quality.A total of 123 healthy subjects were enrolled, and 292 scans were used for the test-retest task.