Mapping the functional connectome between grey matter and white matter Authors: Junchen Zhou, Wenxia Li, Siyuan Luo, Kangjia Chen, Shuo Xu, Qingjin Liu, Huafu Chen, Wei Liao, Jiao Li Published: 2026 ABSTRACT AbstractBrain white matter (WM) has traditionally been viewed as a passive conduit for neural transmission. However, evidence of blood oxygen level-dependent (BOLD) signals measured from the WM suggests its active participation in grey matter (GM) functional networks. Using 7-Tesla functional MRI (fMRI) data, we constructed a GM–WM functional connectome. We found that GM–WM functional architecture follows the unimodal–transmodal hierarchy of GM and is shaped by distributions of neurotransmitter receptors. Distinct WM networks exhibit unique connectivity profiles with GM, reflecting their roles in specific cognitive domains. Individual variations in this connectome correlated with cognitive performance. Notably, compared with the traditional GM–GM functional connectome, the GM–WM functional connectome shows stronger associations with brain disorders, suggesting greater diagnostic sensitivity as a neuromarker. These findings are replicated in a 3-Tesla fMRI cohort. Our work establishes WM as an integral component of the brain’s functional architecture, contributing to hierarchical architecture and supporting higher-order cognition.Subject terms: Cognitive neuroscience, Network models Introduction The brain can be conceptualized as an assembly of neural networks defined by complex patterns of functional connectivity (FC) both within and among networks1. FC measures the temporal synchronization of neural activity across spatially segregated brain regions from functional magnetic resonance imaging (fMRI), representing a fundamental concept in network neuroscience2,3. This network-based neuroimaging approach has significantly advanced our understanding of large-scale brain functional organization. Besides, it reveals dynamic functional reconfigurations across lifespan development and various pathological processes4–6. However, most FC studies have restricted inference to grey matter (GM) as blood oxygen level-dependent (BOLD) signals captured by fMRI predominantly reflect postsynaptic potentials in neuronal dendrites and somata, largely neglecting the potential contributions of white matter (WM)7,8. Recent advances in fMRI acquisition methods have enhanced the feasibility and reliability of detecting BOLD fluctuations within WM8–13. Moreover, such studies have revealed that WM FC is associated with inter-individual differences in cognition14,15 and patterns of WM FC are uniquely altered across diverse neurological and psychiatric disorders, suggesting potential clinical application as diagnostic biomarkers16–18. These observations motivate further studies to integrate WM into functional networks19,20. Emerging evidence suggests that WM dynamically engages in functional circuits during cognitive processing21 and actively modulates the reconfiguration of GM functional networks22. Furthermore, alterations in the GM–WM functional connectome have been consistently identified across various neurological and psychiatric disorders23–26. Thus, comprehensive mapping of the GM–WM functional connectome is essential for understanding neuroprocessing mechanisms and higher-order functions in health and disease. In the current study, we employed high-resolution 7-Tesla resting-state fMRI (rs-fMRI) data from the Human Connectome Project (HCP) to construct the functional connectome between GM and WM (GM–WM FC). First, we mapped the hubs of GM–WM FC and investigated the associated chemo-architecture using positron emission tomography (PET)-estimated neurotransmitter receptor and transporter density data. Second, using the NeuroSynth decoder, we characterized distinct GM–WM FC profiles associated with each WM functional network, revealing the specific contributions of each to cognitive functional specialization, and further investigated the associations between GM–WM FC and the GM functional hierarchy. We also examined the intra-individual reproducibility and predictive value of GM–WM FC as a neuroimaging biomarker of general fluid intelligence (Gf) and for characterizing different clinical cohorts. Results GM–WM FC We reconstructed the functional connectome between GM and WM in healthy young adults (n = 102, age = 29.3 ± 3.1 years, 43 males and 59 females; Supplementary Table S1). The GM was parcellated into 400 parcels according to the Schaefer-400 atlas27. And the 400 GM parcels were assigned to Yeo-7 GM functional networks28 (Fig. 1a): the default mode network (DMN), frontoparietal network (FPN), limbic network (LIN), dorsal attention network (DAN), ventral attention network (VAN), somatosensory network (SMN), and visual network (VN). The WM was parcellated into 200 parcels according to the MWMA-200 atlas29. And the 200 WM parcels were assigned to 7 WM functional networks according to a WM track-weighted dynamic FC study30 (Fig. 1b): the DMN, FPN, SMN, VN, auditory network (AN), brainstem network (BSN), and cerebellum network (CBN). Functional connectivity (FC) between grey matter (GM) and white matter (WM). a The GM was parcellated into 400 parcels, and each parcel was assigned to Yeo-7 GM functional networks28. b The WM was parcellated into 200 parcels, and each parcel was assigned to 7 WM functional networks30. c The whole-brain FC matrix (600 regions × 600 regions) consists of three parts: GM–GM FC matrix (400 regions × 400 regions), WM–WM FC matrix (200 regions × 200 regions), and GM–WM FC matrix (400 regions × 200 regions). d The GM–GM FC exhibited the strongest FC strength [FCGM–GM (r-value) = 0.480 ± 0.144]. The FC strength of WM–WM FC [FCWM–WM(r-value) = 0.174 ± 0.085] was significantly higher than that between GM and WM regions [FCGM–WM(r-value) = 0.161 ± 0.099; two-sample t-test, P < 0.05, Cohen’s d = 0.15]. e Scatter plot showing how FC strength changes with Euclidean distance between regional centroids. All three FC types linearly decreased with Euclidean distance, but this relationship was weakest for GM–WM FC (rGM–GM = − 0.31, P < 0.05; rWM–WM = − 0.22, P < 0.05; rGM–WM = − 0.04, P < 0.05). The whole-brain GM–WM FC matrix (400 regions × 200 regions) revealed extensive connectivity between GM and WM (Fig. 1c). The FC strengths were significantly stronger within WM (Fig. 1d; FCWM–WM = 0.174 ± 0.085) than those with GM (Fig. 1d; FCGM–WM = 0.161 ± 0.099; two-sample t-test, P < 0.05, Cohen’s d = 0.15). Consistent with previous studies15,31, both GM–GM and WM–WM FC strengths decreased with Euclidean distance (Fig. 1e; rGM–GM = − 0.31, rWM–WM = − 0.22; all P < 0.05), whereas the GM–WM FC strength was much less affected by distance (Fig. 1e; rGM–WM = − 0.04, P < 0.05). Chemoarchitecture of GM–WM FC The connectivity pattern of the GM–WM FC was summarized as a weighted degree by summing FC values across WM regions (Fig. 2a), reflecting how strongly GM regions are functionally connected to the whole WM network. The weighted degree pattern followed the hierarchical unimodal–transmodal gradient32 (Supplementary Fig. S1), with the strongest connectivity with the GM-VN and weakest with the GM-LIN and GM-DMN (Fig. 2a). Given the established role of neurotransmitter receptors and transporters in modulating FC patterns17,33,34, we investigated the chemoarchitecture underlying the GM–WM FC matrix. Specifically, we constructed a multiple linear regression model to evaluate the spatial relationship between neurotransmitter systems and the weighted degree profile (Fig. 2b). The model accounted for 24% of the variance in weighted degree pattern (Fig. 2c), with the serotonin and GABA systems showing the highest relative contributions at 37% and 26%, respectively (Fig. 2d). The normality of model residuals supported the model’s validity (Supplementary Fig. S2), and detailed regression coefficients are provided in Supplementary Table S2. The weighted degree pattern of functional connectivity (FC) is related to brain chemo-architecture. a The weighted degree of FC between grey matter (GM) and white matter (WM) was computed by summing the GM–WM FC matrix across all 200 WM regions. The weighted degree was assigned to Yeo-7 GM functional networks and was sorted by the median weighted degree value within each network. In the boxplots, the error bars represent the 10th to 90th percentile range. b A multiple regression model was fitted between eight neurotransmitter system density profiles and the weighted degree pattern. c The regression model explained 24% of the variance in the weighted degree pattern (Adjusted R2 = 0.24). d Using bootstrapping, the relative contribution of each neurotransmitter system was assessed. The serotonin and GABA systems made the highest relative contributions (37% and 26%, respectively). In the median plots, the error bars represent the 5th to 95th percentile range. Functional specialization of WM functional networks The WM exhibits a modular organization comprising multiple functionally distinct networks19,30. To investigate the role of each WM functional network in supporting GM functional activity, we calculated the weighted degree pattern for each WM network to GM (Fig. 3a and Supplementary Fig. S3). The whole-WM weighted degree pattern represented a dominant GM–WM FC profile, as evidenced by high similarity between the weighted degree and each column of the GM–WM FC matrix (Supplementary Fig. S3a; mean Pearson’s r = 0.71 ± 0.03). Therefore, to isolate network-specific contributions, we regressed out the whole-WM weighted degree from the GM–WM FC matrix before calculating the network-level weighted degree pattern (Supplementary Fig. S3b). Each WM network exhibited a distinct connectivity profile with GM (Fig. 3a and Supplementary Fig. S3c). Specifically, the WM-DMN demonstrated strong connections to the GM-DMN. The WM-FPN was strongly coupled to the GM-FPN. The WM–SMN was functionally connected to the GM–SMN. The WM-AN showed strong coupling to the GM-LIN. The WM-VN was most strongly associated with the GM-VN. Lastly, the WM-BSN and WM-CBN were functionally connected to the GM-VAN and GM-FPN. Collectively, these network-specific connectivity profiles suggest that WM functional networks contribute to specialized cognitive processes. Then, for each WM network, we binarized the weighted degree pattern using a 5% threshold to create a mask for NeuroSynth decoding35. We employed the 24 topic terms (v3) as described previously32,36. In Fig. 3b, we presented the five most strongly associated cognitive terms (all PFDR < 0.05) for each WM network. Based on NeuroSynth annotations, the WM-DMN and WM-FPN predominantly support memory-related and cognitive control functions. The WM-SMN is strongly linked to motor processing. The WM-AN is associated with memory-related functions. The WM-VN shows the strongest correlation with visual processing. The WM-BSN supports the pain and inhibition, while the WM-CBN is implicated in memory-related processing. Detailed decoding results for all 24 topic terms are provided in Supplementary Table S3 and S4. Additionally, we comprehensively characterized the cognitive domain spectrum for each WM network by decoding binarized maps across thresholds ranging from 0–5% to 95–100% (Supplementary Fig. S4). The grey matter–white matter functional connectivity (GM–WM FC) pattern suggests widespread roles in cognition, memory, movement, sensory, and executive function. a The network-level weighted degree patterns were calculated by summing the FC of each GM region within the 7 WM functional networks [default mode network (DMN), frontoparietal network (FPN), somatosensory network (SMN), visual network (VN), auditory network (AN), brainstem network (BSN), and cerebellum network (CBN)]. These patterns represent how each WM network is functionally connected with the GM. b Each network-level weighted degree pattern was binarized using a 5% threshold to generate a NeuroSynth-decoding mask. The five most associated terms are shown (all PFDR < 0.05). The detailed decoding results can be found in Supplementary Table S3 and S4. Hierarchical organization of GM–WM FC The connectivity-based gradient represents a continuous spectrum of functional processing that spans across GM regions37. To characterize the hierarchical organization of GM–WM FC, we generated the GM–WM functional gradient from the GM–WM FC matrix (Fig. 4a) and used the traditional GM–GM functional gradient as a reference for the intrinsic GM hierarchy (Supplementary Fig. S5). A strong spatial correspondence was observed between the GM–WM gradient and GM–GM gradient (Fig. 4b; r = 0.83, PMoran < 0.05), confirming that the GM–WM FC reflects GM hierarchical organization. However, these distributions diverged significantly, with the GM–WM gradient exhibiting a centralized distribution, suggesting functional convergence among WM-connected GM regions (Fig. 4c), while the GM–GM gradient displayed greater dispersion, reflecting broader functional differentiation (Fig. 4c). To further quantify these topological differences, we computed global, between-network, and within-network dispersion metrics38 (Supplementary Fig. S6). At the global level, the GM–WM gradient exhibited significantly lower dispersion than the GM–GM gradient (Fig. 4d; paired t-test, P < 0.05, Cohen’s d = −0.89), indicating a more compact whole-brain organization. The GM–WM gradient also showed significantly reduced between-network dispersion (Fig. 4d; paired t-test, P < 0.05, Cohen’s d = −0.94), reflecting less functional differentiation across networks in the low-dimensional space when considering WM rs-fMRI signals. Divergent patterns were observed for within-network dispersion. Within higher-order associative networks, including the DMN, FPN, and LIN, the GM–WM gradient displayed significantly lower within-network dispersion compared with the GM–GM gradient (Fig. 4d; all PFDR < 0.05, Cohen’s dDMN = −0.38, Cohen’s dFPN = −0.88, Cohen’s dLIN = −1.37), suggesting greater functional homogeneity. In contrast, within primary networks, including the VN and SMN, the GM–WM gradient displayed higher within-network dispersion (Fig. 4d; all PFDR < 0.05, Cohen’s dVN = 1.00, Cohen’s dSMN = 1.46), indicating greater functional heterogeneity. Hierarchical organization of grey matter–white matter functional connectivity (GM–WM FC). a The GM–WM gradient was calculated from the GM–WM FC. The GM–WM FC was thresholded by retaining only the strongest 10% of connections for each row. An affinity matrix was then constructed by computing the cosine similarity between any two rows of the thresholded GM–WM FC matrix. Finally, the principal gradient was obtained by diffusion map embedding. b We observed a strong spatial correlation between the GM–WM gradient and GM–GM gradient (r = 0.83, PMoran < 0.05). c The GM–GM gradient showed a dispersed distribution, whereas the GM–WM gradient showed a centralized distribution. d At the global level, the GM–WM gradient demonstrated a significantly lower dispersion than the GM–GM gradient. The GM–WM gradient showed a significantly lower dispersion than the GM–GM gradient at the between-network level. At the within-network level, the GM–WM gradient also showed significantly lower dispersion in default mode network (DMN), frontoparietal network (FPN), and limbic network (LIN), and greater dispersion in visual network (VN), and somatosensory network (SMN). In the boxplots, the error bars represent the 10th to 90th percentile range. Individual variability of GM–WM FC The GM–WM FC patterns were highly variable among subjects (Fig. 5a; mean Pearson’s r = 0.47 ± 0.11). To investigate the functional implications of this individual variability, we assessed the capacity of the GM–WM FC pattern to identify individuals (identifiability) and predict cognitive performance39,40. First, to construct an identifiability matrix, pairwise Pearson’s correlation coefficients were computed between FC profiles of subjects across the HCP 7-Tesla test and retest cohorts (Fig. 5b), and then the maximal correlation value for each row (test-retest identification) and column (retest-test identification) was selected for individual identification. This process yielded 55% test-retest accuracy (56/102) and 57% retest-test accuracy (58/102; Fig. 5b), and permutation tests confirmed the statistical significance of these identification rates [maximal chance accuracy: 5/102 (5%), thus Pperm < 0.05]. However, the identification accuracy of the GM–WM FC matrix was lower than that of the GM–GM FC matrix (Supplementary Fig. S7a; 89% for test-retest identification, 92% for retest-test identification). This reduction likely stems from the lower temporal signal-to-noise ratio (tSNR) of WM rs-fMRI signals (Supplementary Fig. S8; two-sample t-test, P < 0.05, Cohen’s d = −0.43), which compromises the reliability of FC estimates. Individual variation in grey matter–white matter functional connectivity (GM–WM FC) is associated with general fluid intelligence (Gf). a The GM–WM FC matrices from two participants are shown to demonstrate individual variation in GM–WM FC. Pearson’s correlation coefficients were calculated between individual GM–WM FC matrices and the group-average GM–WM FC matrix to quantify individual variability. b The identifiability matrix was constructed by computing the Pearson’s correlation coefficients for the vectorized GM–WM FC matrices of all subject pairs between the HCP 7-Tesla test and retest cohorts. The diagonal elements of the identifiability matrix represent the within-subject similarity, while the non-diagonal elements represent between-subjects similarity. The test–retest and retest–test identification accuracies were 55% and 57%, respectively. The maximal chance accuracy (5%) is marked by a horizontal line in the bar graph for reference. c The GM–WM FC-based model predicted Gf scores in both the HCP 7-Tesla test cohort (r = 0.30, Pperm < 0.05) and retest cohort (r = 0.17, Pperm < 0.05). Dashed lines indicate the 95% confidence band of the linear regression fit. Given the established individual uniqueness of GM–WM FC, we next investigated its utility as a neuromarker of cognitive ability. A connectome-based predictive model trained on the HCP 7‑Tesla test cohort significantly predicted Gf scores within the same cohort (Fig. 5c; r = 0.30, Pperm < 0.05) and generalized to the retest cohort (Fig. 5c; r = 0.17, Pperm < 0.05). By contrast, the predictive model based on GM–GM FC showed weaker within-cohort prediction (Supplementary Fig. S7b; r = 0.21, Pperm < 0.05) and failed to generalize to the retest cohort (Supplementary Fig. S7b; r = -0.03, Pperm = 0.44). The stronger predictive performance and generalizability of the GM–WM FC model suggest that the neural substrates supporting higher-order intellectual functions extend beyond GM networks to include WM networks14. Brain disorders classification performance To assess the discriminative power of GM–WM FC for brain disorders, we enrolled three clinical cohorts comprising Alzheimer’s disease (AD), autism spectrum disorder (ASD), and major depressive disorder (MDD) patients. We attempted to distinguish patients from corresponding healthy controls (HCs). To mitigate overfitting due to the high dimensionality of the FC, a least absolute shrinkage and selection operator (LASSO) was embedded in the classification pipeline for feature selection. The selected features from the GM–GM FC and GM–WM FC matrices were then used separately as input to a support vector machine (SVM) classifier41. Detailed demographic information of brain disorders cohorts is provided in Supplementary Table S5. Across three brain disorder cohorts, the GM–WM FC matrix consistently outperformed the GM–GM FC matrix in distinguishing patients from HCs, suggesting that GM–WM FC may be a more sensitive neuromarker for these disorders. Specifically, for AD vs. HC classification, GM–WM FC achieved a balanced accuracy of 0.78 (Pperm < 0.05), compared to 0.73 (Pperm < 0.05) of GM–GM FC. A similar trend was observed in the ASD vs. HC classification, with GM–WM FC achieving balanced accuracy of 0.85 (Pperm < 0.05), whereas GM–GM FC attained lower accuracy of 0.74 (Pperm < 0.05). For MDD vs. HC classification, both networks exhibited limited discriminative ability. GM–WM FC attained a balanced accuracy of 0.66 (Pperm < 0.05), whereas GM–GM FC performed near chance level (balanced accuracy = 0.53, Pperm = 0.25). Complete classification performance metrics are provided in Supplementary Table S6. Validation results For validation, we first evaluated the reproducibility of weighted degree patterns across a range of connection densities (10%–90% in 10% increments) and found high consistency at all densities (Supplementary Fig. S9). Second, the GM–WM FC matrix and the weighted degree pattern were recalculated from the HCP 7-Tesla retest cohort, which revealed both excellent test-retest reliability (TRR) for the GM–WM FC matrix (Supplementary Fig. S10a; r = 0.95, P < 0.05) and high spatial consistency of the weighted degree pattern (Supplementary Fig. S10b; r = 0.97, PMoran < 0.05). In addition, there were significant correlations between FC matrices derived from 3-Tesla and 7-Tesla cohorts (Supplementary Fig. S10a; r = 0.55, P < 0.05) and strong correspondence of weighted degree patterns across field strengths (Supplementary Fig. S10b; r = 0.88, PMoran < 0.05). Third, we reconstructed the GM–WM FC matrix using the Schaefer-200 atlas (Supplementary Fig. S11a) and confirmed that the weighted degree pattern preserved its VN-to-LIN spatial distribution (Supplementary Fig. S11b). Fourth, we replicated the patterns of the GM–WM FC matrix (Supplementary Fig. S12a; r = 0.98, P < 0.05) and corresponding weighted degree (Supplementary Fig. S12b; r = 0.98, PMoran < 0.05) using the 7-Tesla data without applying spatial smoothing in the preprocessing pipeline. Next, we replicated the patterns of GM–WM FC matrix (Fig. 6a; r = 0.56, P < 0.05) and weighted degree across a range of connection densities [Fig. 6b; 1%–10% in 1% increments, r values ranging from 0.20 to 0.54, all PMoran < 0.05, false discovery rate (FDR) corrected] using data with global signal regression (GSR). To provide a more comprehensive assessment of GSR sensitivity, we extended this analysis to thresholds from 11% to 20% (Supplementary Table S7). Spatial correlations between weighted degree maps with and without GSR became non-significant (PMoran > 0.05) starting at the 12% threshold and above, indicating that the influence of GSR on network topography emerges at higher connection densities. Besides, the TRR of GM–WM FC was higher for data processing without GSR than with GSR (Supplementary Fig. S13a; TRRwith GSR = 0.27 ± 0.13, TRRwithout GSR = 0.35 ± 0.13, paired t-test, P < 0.05, Cohen’s d = -0.63), whereas performing GSR had a negligible effect on the reliability of GM-GM FC (Supplementary Fig. S13b; TRRwith GSR = 0.69 ± 0.11, TRRwithout GSR = 0.70 ± 0.09, paired t-test, P = 0.68, Cohen’s d = -0.03). Finally, we observed a moderate spatial correlation between the tSNR map and the weighted degree pattern (Supplementary Fig. S14a and b; r = 0.46, PMoran < 0.05). We also compared the weighted degree pattern and the tSNR map (both z-scored) using a paired t-test. We found widespread significant regional differences, suggesting that the weighted degree pattern of GM–WM FC is not simply an artifact of tSNR variation (Supplementary Fig. S14b). Replication of grey matter–white matter functional connectivity (GM–WM FC) with global signal regression (GSR). a We replicated the pattern of GM–WM FC using data preprocessing with GSR. This application of GSR reduced GM–WM FC strength and introduce more negative edges in the FC matrix. b The weighted degree pattern of GM–WM FC was also replicated using across a range of connection densities (1-10% in 1% increments). All PMoran values were adjusted by false rate discovery (FDR). Discussion In this study, we investigated the functional interactions between GM and WM using high-resolution 7-Tesla rs-fMRI data. We found that GM–WM FC follows a spatial distribution aligned with the unimodal-transmodal hierarchical axis, and is shaped by the spatial distributions of neurotransmitter receptors and transporters. Furthermore, individuals’ GM–WM FC patterns were sufficiently unique for identification and were predictive of general cognitive capacity. Notably, GM–WM FC patterns distinguished patients with brain disorders from HCs more accurately than traditional GM–GM FC patterns, suggesting that GM–WM FC may serve as a more sensitive diagnostic neuromarker. Elucidating the fundamental principles of brain functional organization remains a central challenge in neuroscience. Although advances in rs-fMRI have facilitated the characterization of WM FC15,30, functional interactions between GM and WM are not well described19–21. Mounting evidence indicates that GM–WM functional interactions manifest across multiple spatial and temporal scales22. Notably, a recent 7-Tesla rs-fMRI study reported that brainstem activity modulates GM function at multiple temporal scales42, underscoring the importance of investigating subcortical–GM interactions. We extended these findings by comprehensively mapping interactions between WM and GM at the whole-brain level using a recently developed multimodal WM atlas29 to provide an integrated framework for understanding how WM networks modulate GM activity. The distributions of serotonin and GABA receptors and transporters accounted for a substantial proportion of the spatial variability in GM–WM FC, suggesting that these systems are major modulators of GM–WM FC patterns. Previous studies have reported that regional GABA concentrations are correlated with FC strength43,44, and serotonin influences intrinsic FC dynamics45, underscoring the critical involvement of these neurotransmitter systems in shaping the structure of large-scale functional networks46. Moreover, serotonin and GABA may act synergistically to regulate functional brain networks by dynamically fine-tuning the global excitation–inhibition balance47,48. Future studies should examine whether disease-related alterations in GM–WM FC are associated with dysregulation of these neurotransmitter systems, thereby providing novel molecular clues to the pathophysiology of brain disorders. Beyond the dominant connectivity pattern, each WM functional network exhibited distinct connectivity profiles with GM regions, reflecting widespread involvement in diverse cognitive and behavioral processes. Specifically, the WM-DMN demonstrated strong associations with memory-related and cognitive-related functions, consistent with the well-established role of the GM-DMN in supporting self-referential processing49,50. This parallel functional specialization underscores the tight coupling between corresponding functional networks in WM and GM19. The WM-FPN showed strong connectivity with GM-FPN regions involved in working memory and cognitive control, consistent with recent tract-based lesion network symptom association studies51. These observations highlight the critical role of both anatomical WM tracts and functional WM networks in supporting higher-order executive functions52. Primary sensory networks displayed specialized connectivity patterns, with the WM-SMN predominantly linked to motor functions, while the WM-VN and WM-AN were selectively associated with visual perception-related and semantic functions42. These findings provide compelling evidence for functional specialization within WM networks. Further, the WM-CBN and WM-BSN were primarily implicated in inhibition-related and emotion-related functions, suggesting unique contributions to emotional regulation and cognitive control processes53. Collectively, these results demonstrate that specialized WM functional networks support distinct cognitive domains. The GM–WM functional gradient spatially aligns with the well-known unimodal–transmodal GM gradient axis32,54. Traditionally, characterization of this gradient has relied solely on GM–GM FC, overlooking the potential contributions of WM. Our findings indicate that the functional architecture captured by the GM–WM gradient is distinct from that revealed by GM–GM FC, exhibiting overall reduced functional differentiation and network-specific patterns of functional homogeneity. The lower global and between-network dispersion of the GM–WM gradient suggests that GM–WM FC may reflect a more integrative underlying infrastructure. While GM–GM FC maps the specialized, modular organization of functional networks, GM–WM FC likely captures aspects of their shared organization. The divergent within-network dispersion patterns provide a more nuanced interpretation. The lower dispersion within higher-order networks such as the GM-DMN and GM-FPN via GM–WM FC may indicate that these distributed networks are supported by strong, coherent functional coupling with WM pathways. This may reflect the global hub role of higher-order networks in accommodating the diverse demands of complex cognitive functions49,55–57. In contrast, the higher dispersion within primary networks such as GM-SMN and GM-VN indicates that these networks possess more localized functional topology, which likely facilitates the rapid processing of unimodal information within cortical regions32,58. Future investigations should explore developmental38, aging-related6, and clinical variations59 in GM–WM gradient patterns to gain a deeper understanding of brain organization across the lifespan and under pathological conditions. The GM–GM FC has been established as a unique neural ‘fingerprint’ capable of distinguishing individuals39,60. To assess whether GM–WM FC exhibits similar discriminative power, we analyzed test-retest data from the HCP 7-Tesla cohort and found significant identification accuracy, albeit lower than that of GM–GM FC (Supplementary Fig. S7a). This reduced identification accuracy may stem from the inherently lower tSNR of BOLD signals from WM compared to GM (Supplementary Fig. S8). To address this limitation, future research should focus on developing WM-optimized functional imaging sequences and tailored preprocessing pipelines to improve the specificity and tSNR of WM functional signals7,8,61. One promising advance in this direction is apparent diffusion coefficient fMRI (ADC-fMRI), a technique that quantifies neural activity-induced microstructural changes by measuring water molecule diffusion dynamics62. The ADC-fMRI may offer enhanced sensitivity and specificity for detecting WM neural activity and elucidating WM functional architecture. Our results also demonstrate that impacts of GSR in GM–WM FC. While data processing with GSR effectively reduces artifacts in standard GM–GM FC studies, we found that it decreased the TRR of GM–WM FC, leaving GM–GM FC reliability unaffected (Supplementary Fig. S13), possibly suggesting that global signal contains neurophysiologically relevant information shared between GM and WM. Given that GSR choice can substantially alter observed FC, future studies of GM–WM interactions should carefully consider its inclusion in the preprocessing pipeline. Growing evidence suggests that aberrant functional coupling between GM and WM represents a characteristic neuropathological feature across various brain disorders23,24. Our findings further substantiate this view by demonstrating the significantly greater sensitivity of GM–WM FC patterns as neuroimaging biomarkers for clinical classification of AD, ASD, and MDD compared to conventional GM–GM FC measures. This enhanced discriminative capacity likely stems from two key factors. First, GM–WM functional interactions appear particularly vulnerable to disease-related disruptions. Second, the integrated analysis of GM and WM signals provides a more comprehensive characterization of neural network dysfunction, capturing pathological changes that may go undetected when GM is examined in isolation. Notably, while our current investigation focused on static FC metrics, emerging research indicates that brain functional organization is fundamentally dynamic in nature63,64. Thus, static GM–WM FC measurements may mask crucial temporal variations. Future studies should examine whether dynamic FC metrics demonstrate greater clinical utility than static measures for revealing disease-specific alterations in GM–WM interactions65,66. Several limitations should be addressed. First, we analyzed data from only young adults, whereas the dynamic nature of brain network maturation suggests the need for lifespan studies. In addition, the sample size remains a limitation for the generalizability of our findings. Future studies with larger samples are warranted to confirm the generalizability of the GM–WM connectome architecture and to examine its associations with individual differences in demographics, genetics, and behavior. Second, the current analysis was limited by the use of rs-fMRI and neurotransmitter density data from different cohorts. Subsequent studies should employ multimodal imaging (e.g., rs-fMRI and PET) in the same participants to enable precise characterization of the neurochemical substrates underlying GM–WM FC patterns and individual-level analyses of structure-function-neurotransmitter relationships. Third, although we established a significant relationship between GM–WM FC and Gf, this metric represents only one dimension of higher-order cognition. Expanding the assessment to include additional cognitive domains would provide a more comprehensive understanding of how GM–WM interactions support diverse cognitive processes. Finally, although GM–WM FC demonstrated higher accuracy than GM–GM FC in distinguishing the three clinical cohorts from their respective HCs in this preliminary investigation, these findings should be interpreted with caution. The present analyses were conducted in relatively small clinical samples, with incomplete case-control matching, particularly in the ASD cohort, which showed a notable imbalance in sex ratios between patients and HCs. An important future direction is to validate the discriminative capacity of GM–WM FC in larger, well-matched clinical cohorts with balanced demographic characteristics, to rigorously evaluate its potential as a reproducible neuroimaging biomarker for psychiatric and neurological disorders. This study reveals a complex functional connectome between GM and WM that contributes to the hierarchical organization of the GM, supports widespread cognitive processing, and serves as a robust neuroimaging biomarker for classifying brain disorders. Collectively, these findings provide a novel framework for understanding human brain functional architecture, complementing and extending traditional GM-centric models by highlighting the essential role of WM in large-scale brain network organization and cognitive function. Methods Participants and MRI data acquisition HCP 7-Tesla test-retest cohort All four 16-minute (900 TRs) rs-fMRI scans (two scans per day) from the HCP S1200 release 7-Tesla cohort67, namely REST1_PA, REST1_AP, REST2_PA, and REST2_AP, were included in the analysis (where ‘PA’ and ‘AP’ indicate posterior-to-anterior and anterior-to-posterior phase coding directions). Scans REST1_PA and REST1_AP were concatenated as the 7-Tesla test cohort, while REST2_PA and REST2_AP were concatenated as the 7-Tesla retest cohort. All rs-fMRI images were acquired on a 7-Tesla scanner (Siemens Healthcare, Germany) using a gradient-echo echo planar imaging (EPI) sequence at 1.6 mm isotropic resolution (Supplementary Table S8). Of 184 participants who underwent scanning, 175 participants completed all four rs-fMRI scans. Following quality control, 73 subjects were excluded due to a root-mean-square head position change > 0.8 mm (half a voxel width) across the four rs-fMRI scans68, resulting in a final analytic cohort of 102 healthy young adults. Detailed demographics are provided in Supplementary Table S1. Participant recruitment procedures and informed consent forms, including consent to share de-identified data, were approved by the Washington University Institutional Review Board as part of the HCP. HCP 3-Tesla cohort Only the 14-min (1200 TRs) REST1_LR scan from the HCP S1200 3-Tesla unrelated cohort was analyzed as the replication cohort. The rs-fMRI images were acquired on a customized 3-Tesla Connectome Skyra scanner (Siemens Healthcare, Germany) using a gradient-echo EPI sequence at an isotropic resolution of 2 mm (Supplementary Table S8). More details about the imaging parameters can be found in a previous study67. Data from all 100 participants were included. Detailed demographics are provided in Supplementary Table S1. Clinical cohorts We also conducted analyses on public datasets of AD patients (n = 34, rs-fMRI and T1w data; ADNI 2; https://adni.loni.usc.edu/data-samples/adni-data), ASD patients (n = 62, rs-fMRI and T1w data; ABIDE I; https://fcon_1000.projects.nitrc.org/indi/abide/), and MDD patients (n = 50, rs-fMRI and T1w data; SPRBS; https://www.synapse.org/Synapse:syn22317081). These clinical datasets also included the rs-fMRI and T1w scans from 46, 94, and 41 demographically matched HCs, respectively. The detailed acquisition parameters are listed in Supplementary Table S9, and detailed demographics are listed in Supplementary Table S5. GM and WM Parcellations construction The Schaefer-400 atlas27 was used for GM parcellation and the MWMA-200 atlas29 for WM parcellation in the primary analysis (Fig. 1a and b). The 400 GM parcels were assigned to Yeo-7 functional networks28, the DMN, FPN, LIN, DAN, VAN, SMN, and VN (Fig. 1a). The 200 WM parcels were assigned to 7 WM functional networks (Fig. 1b; DMN, FPN, SMN, VN, AN, BSN, and CBN) based on a previous study30. In that work, ten WM independent components (ICs) were derived from tract-weighted dynamic FC, including IC1-Medial Sensorimotor, IC2-Medial Visual, IC3-Default Mode, IC4-Auditory, IC5-Lateral Sensorimotor, IC6/7-Frontoparietal, IC8-Premotor, IC9-Lateral Visual, IC10-Vertical Occipital. The ten ICs were first merged into five WM networks: DMN (IC3), FPN (IC6 and IC7), SMN (IC1, IC5, and IC8), VN (IC2, IC9, and IC10), and AN (IC4). In addition, the CBN and BSN were included in the MWMA-200 atlas29. In the replication analysis, we used the Schaefer-200 atlas for GM parcellation. The Schaefer and MWMA atlases were also used as GM and WM masks, respectively, for the spatial smoothing of rs-fMRI data during preprocessing. rs-fMRI data preprocessing HCP 7-Tesla test–retest cohort The rs-fMRI images from the HCP 7-Tesla cohort were first preprocessed through the HCP fMRI Volume pipeline. Further preprocessing included the following steps. First, spatial normalization was performed using the Montreal Neurological Institute (MNI) template with a 3 × 3 × 3 mm3 voxel size to enhance computational efficiency59. Second, 12 rigid-body motion parameters and mean CSF signals were regressed out of the rs-fMRI time series69. To avoid elimination of important neural signals, the GSR was not performed8,15,19. Then, time series were band-pass filtered at 0.01 − 0.1 Hz to eliminate unphysiological signals2. The spatial smoothing was performed separately within the GM and WM masks using a 6-mm full-width at half maximum (FWHM) Gaussian kernel (Supplementary Fig. S15), ensuring that no cross-tissue blurring occurred between GM and WM14–16. The four rs-fMRI time series were then demeaned, and the REST1_PA and REST1_AP time series were concatenated as the 7-Tesla test cohort, while the REST2_PA and REST2_AP time series were concatenated as the 7-Tesla retest cohort. Thus, both the test and retest 7-Tesla rs-fMRI time series consisted of 1800 TRs. HCP 3-Tesla cohort The rs-fMRI images from the HCP 3-Tesla cohort were first preprocessed through the HCP fMRI Volume pipeline. Subsequent preprocessing steps were applied uniformly to both the 7-Tesla and 3-Tesla cohorts, including spatial normalization, regression of 12 rigid-body motion parameters and the mean CSF signal, band-pass filtering, and spatial smoothing separately within the GM and WM masks. The 3-Tesla rs-fMRI time series consisted of 1200 TRs. Clinical cohorts The preprocessing protocol was applied uniformly across AD, ASD, and MDD clinical cohorts. All functional images were preprocessed using the DPARSF toolkit70. Preprocessing involved elimination of the first 10 volumes, slice-timing correction, realignment to the functional volumes, co-registration with structural images, multiple regression of 24 rigid-body motion parameters, and regression of mean CSF signal from the functional images. Preprocessed images were band-pass filtered at 0.01 − 0.1 Hz, normalized to MNI space, resampled to 4 × 4 × 4 mm3 voxel size, and spatially smoothed using an 8 mm FWHM Gaussian kernel separately within GM and WM masks. The lengths of the rs-fMRI time series in AD, ASD, and MDD cohorts were 130, 170, and 133 TRs, respectively. Reconstruction of the GM–WM functional connectome A GM–WM FC matrix (400 regions × 200 regions) was constructed for each participant by calculating Pearson’s correlation coefficients between the averaged rs-fMRI time series of all GM–WM pairs. Matrices were then normalized using Fisher’s r-to-z transformation. Individual z matrices were averaged across subjects to generate a group-average GM–WM FC matrix, and inverse Fisher’s r-to-z transformation was performed. The GM–WM FC pattern was then summarized as the weighted degree by adding group-average GM–WM FC values across WM regions. To investigate the functional specialization of WM functional networks, we first regressed out the whole-WM weighted degree from the GM–WM FC matrix to obtain the GM–WM regressed-FC matrix42 (Supplementary Fig. S3b). The connectivity pattern for each WM functional network was obtained by summing the regressed FC values across WM regions within that network (Supplementary Fig. S3c). Chemoarchitecture of connectivity profile To decode the chemoarchitecture of the GM–WM FC matrix, we obtained density images for 18 neurotransmitter receptors and transporters from neuromaps33. For receptors and transporters with more than one mean image of the same tracer (i.e., 5-HT1b, D2, VAChT, and mGluR5), subject-weighted averaging was performed71. All density images were first down-sampled to a 3 × 3 × 3 mm³ voxel size to match the spatial resolution of the brain atlas used in this study. These density images were then parcellated into 400 GM regions using the Schaefer-400 atlas. Each density image was z-normalized across the 400 regions. Neurotransmitter receptors and transporters from the same neurotransmitter system were averaged to generate 8 system images: serotonin (5-HT1A, 5-HT1B, 5-HT2, 5-HT4, 5-HT6 and 5-HTT), dopamine (D1, D2 and DAT), norepinephrine (NET), histamine (H3), acetylcholine (ɑ4β2, M1 and VAChT), cannabinoid (CB1), glutamate (NMDA and mGluR5) and GABA (GABAA/BZ). A multivariate linear regression model was constructed to explore the contributions of neurotransmitter systems to the GM–WM connectivity pattern (Fig. 2b), and the relaimpo package in R was used to compute the relative importance metrics72. The regression model was defined as where degree is the weighted degree of GM–WM FC. We then assessed the relative contribution of each neurotransmitter system by bootstrapping. NeuroSynth decoder To investigate the contributions of the 7 WM networks to distinct cognitive processes, functional activation maps were obtained from the NeuroSynth meta-analytic database35. In this study, 24 topic terms (v3) were included32,36, including action, auditory processing, autobiographical memory, cognitive control, cued attention, declarative memory, emotion, eye movements, face/affective processing, inhibition, language, motor, multisensory processing, numerical cognition, pain, reading, reward-based decision making, social cognition, verbal semantics, visual attention, visual perception, visual semantics, visuospatial, and working memory (Supplementary Fig. S4 and Table S3). For each WM network, the weighted degree pattern was binned into twenty binarized maps ranging from 0–5% to 95–100%. In the primary analysis, the 95–100% map, representing GM regions with the strongest connectivity to a specific WM network, was used as input to the NeuroSynth decoder71. The decoder outputs a z-statistic for each topic term, which was then converted to a p-value and corrected using FDR correction (Fig. 3). To fully characterize the cognitive domain spectrum associated with each WM network, all twenty binarized maps were decoded in the supplemental analysis (Supplementary Fig. S4). Connectome-based gradients analyses Connectivity-based gradients capture the global topography of brain32,37. We identified the traditional GM–GM functional gradient (Supplementary Fig. S5), which represents the functional hierarchical organization of the GM, and the GM–WM functional gradient (Fig. 4a), which represents how similar GM regions are functionally connected to the WM. The spatial similarity of the GM–WM gradient and GM–GM gradient was computed, and the statistical significance of the correlation result was determined using 5,000 iterations of the Moran spectral randomization (MSR) test73,74. In detail, we identified the principal GM functional gradient for both the group-average GM–WM FC matrix (400 regions × 200 regions) and the group-average GM–GM FC matrix (400 regions × 400 regions) using the BrainSpace toolbox74. The GM–GM FC matrix was obtained by calculating Pearson’s correlations between averaged rs-fMRI time series of paired brain regions within the GM parcellation. The gradient mapping protocol was the same for these two FC matrices. First, the FC matrix was row-wise thresholded by retaining only the top 10% connections of each row75,76. Then, we computed the cosine similarity between any two rows of the thresholded FC matrix to obtain the affinity matrix (400 regions × 400 regions). Finally, diffusion map embedding was implemented to derive the principal gradient. The diffusion map embedding is a nonlinear dimensionality reduction algorithm. The algorithm is controlled by two parameters: ɑ, which controls the influence of the density of sampling points on the underlying manifold, and t, which controls the scale of the diffusion operator’s eigenvalues. We set the ɑ at 0.5 and t at 0, preserving global relations between regions in the embedding space75,77. To compare the topological organization of GM–GM FC and GM–WM FC matrices in the low-dimensional embedding space38,78, we quantified global, between-network, and within-network dispersion. Specifically, global dispersion was defined as the Euclidean distance from each region to the manifold centroid (Supplementary Fig. S6a), providing a whole-brain index of integration and segregation. Between-network dispersion was measured as the Euclidean distance between network centroids, reflecting the differentiation among networks (Supplementary Fig. S6b). Within-network dispersion was computed as the Euclidean distance between each region and its respective network centroid (Supplementary Fig. S6c), capturing the similarity of FC profiles within a given network. Before the computation of dispersion measures, both gradients were rescaled from 0 to 1 to exclude scaling effects. Connectome-based fingerprinting analyses The GM–GM FC pattern can serve as a ‘fingerprint’ that accurately identifies subjects within a large group39,60. In this study, we tested whether the GM–WM FC pattern is also sufficiently unique at the individual level for participant identification. We first constructed an identifiability matrix (102 × 102) by computing Pearson’s correlation coefficients of vectorized GM–WM FC matrices (80,000 × 1) of all subject pairs between the HCP 7-Tesla test cohort and HCP 7-Tesla retest cohort79, yielding diagonal elements representing within-subject similarity and off-diagonal elements representing between-subject similarity. The identification procedure was then performed by selecting the maximum element in each row (test-retest identification) or column (retest-test identification), with identification accuracy defined as the number of correct identifications divided by the number of subjects (102 subjects in this study). Finally, we performed a permutation test to assess the statistical significance of the accuracy by shuffling the identifiability matrix and executing the identification procedure 5,000 times. The Pperm value was defined by dividing all iterations with a number greater than the true accuracy by the total number of iterations. For comparison, the same fingerprinting procedure was performed using the GM–GM FC matrix. Connectome-based prediction analyses If the GM–WM FC pattern is unique to each individual, the next question is whether this pattern can be used as a neuromarker to predict an individual’s behavioral performance. Using the connectome-based prediction model14,40, we assessed the capacity of the individual GM–WM FC pattern to predict the general intelligence factor Gf. In each leave-one-out cross-validation (LOOCV) iteration, we excluded one subject, leaving the remaining subjects as the training set. The significant (P < 0.01) relevant edges to Gf scores in the training set were then selected and separated into positive and negative subsets. The edges in the positive and negative subsets were summed separately to generate positive and negative features. Based on a general linear model, we merged the two features, and this predictive model was then applied to the left-out subject to obtain the predicted Gf score. The LOOCV procedure was repeated to obtain each subject’s predicted Gf score, and Pearson’s correlation coefficient was calculated between observed and predicted Gf. To obtain the Pperm value of correlation, a permutation test was performed by shuffling the observed Gf scores 5000 times and executing the LOOCV procedure to generate the null distribution of correlation coefficient. The Pperm value was defined by dividing all iterations with a number greater than the true correlation coefficient by the total number of iterations. To ensure the generalizability of the prediction model, we trained the model using the 7-Tesla test cohort and tested the model using both the 7-Tesla test cohort and the 7-Tesla retest cohort. For comparison, the same prediction procedure was performed using the GM–GM FC matrix. Brain disorders classification Previous studies have documented alterations in the GM–GM FC matrix among clinical populations with mental illness5,65,71 and recent evidence further suggests that such functional dysregulation also extends to GM–WM interactions. To investigate whether the GM–WM FC pattern may serve as a viable neuroimaging biomarker for psychiatric disorders, we enrolled three disease cohorts with diagnosed AD, ASD, and MDD, and constructed the GM–GM FC and GM–WM FC matrices for each subject. Features were derived by extracting the upper triangular elements of the GM–GM FC matrix and all elements of the GM–WM FC matrix. Dimensionality reduction and feature selection were performed using the LASSO. A SVM classifier41 was then employed to distinguish patients from HCs. We used the LOOCV strategy to evaluate the discriminative ability of the two FC matrices80. In each iteration, one subject was held out as the test sample, and the remaining subjects were used to train the SVM model. Classification performance was quantified using balanced accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC-ROC). To determine the statistical power, permutation testing was conducted for each classification model a total of 100 times. Validation analyses To evaluate the potential influence of thresholding on our findings81, we reconstructed weighted degree patterns of GM–WM FC matrices across a range of connection densities (Supplementary Fig. S9; 10% to 90%, in 10% increments). The spatial consistency between each threshold-derived degree pattern and the primary result (Fig. 2a; obtained without thresholding) was then quantified. We further assessed TRR by deriving the GM–WM FC matrix and weighted degree pattern from the HCP 7-Tesla retest cohort (Supplementary Fig. S10). To rule out potential field-strength dependence61, the same metrics were computed using the unrelated HCP 3-Tesla cohort (Supplementary Fig. S10). Third, we examined the robustness of our findings to atlas resolution by reconstructing the GM–WM FC matrix and weighted degree pattern using the Schaefer-200 atlas (Supplementary Fig. S11). To evaluate the impact of spatial smoothing82,83, we repeated the analysis in the HCP 7-Tesla cohort without applying spatial smoothing (Supplementary Fig. S12). Furthermore, to assess the influence of GSR3,84,85, we recomputed GM–WM FC matrices by performing GSR in the same cohort. Given that performing GSR substantially reduces coupling strength and introduces negative edges (Fig. 6a), we restricted subsequent weighted degree calculations to the strongest connections (Fig. 6b; 1% to 20% density, in 1% increments, applied row-wise). To further evaluate the impact of GSR, we assessed the TRR of GM–WM FC and GM–GM FC with and without GSR (Supplementary Fig. S13). The TRR was defined as the Pearson’s correlation coefficient between FC matrices derived from the test and retest session of the same participants in the HCP 7-Tesla dataset. Finally, we examined the spatial relationship between the weighted degree pattern and the tSNR map of rs-fMRI signals to exclude potential noise-driven effects42,77 (Supplementary Fig. S14). Statistics and reproducibility To account for spatial autocorrelation in MRI data, all spatial correlations between brain maps were assessed using MSR test with 5,000 iterations, and PMoran values were corrected for multiple comparisons by FDR. Permutation tests with 5,000 iterations were used to evaluate the significance of identifiability accuracy, and prediction performance. Permutation test with 100 iterations was used to evaluate the significance of classification performances. For reproducibility, we replicated main findings in the HCP 7-Tesla retest cohort, HCP 3-Tesla cohort, and across different scale atlases (Schaefer-200), preprocessing pipelines (without spatial smoothing, with global signal regression), and a range of connection densities. Test-retest reliability of FC matrices was quantified using Pearson’s correlation coefficient. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Material Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Author contributions J.Z. and J.L. designed the study; J.Z. conducted the primary analyses; S.L. and K.C. conducted validation analyses; S.X. and Q.L. contributed to the visualization and interpretation of the results; J.Z. drafted the manuscript; H.C., W.L., and J.L. critically revised the manuscript. All authors reviewed and approved the final version. Peer review Peer review information Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editor: Jasmine Pan. A peer review file is available. Funding W.L. discloses support for the research of this work from the National Key R&D Program of China (2022YFC2009906), the National Natural Science Foundation of China (62571105), and the Sichuan Science and Technology Program (2026NSFSCZY0055). J.L. discloses support for the research of this work from the National Natural Science Foundation of China (62473082), the Sichuan Science and Technology Program (2026NSFSC0377), and the Fundamental Research Funds for the Central Universities (ZYGX2024XJ054). H.C. discloses support for publication of this work from the National Natural Science Foundation of China (62036003). Data availability Structural and functional neuroimaging data of healthy participants were obtained through the Washington University-University of Minnesota (WU-Minn) Human Connectome Project (https://www.humanconnectome.org/software/connectomedb). PET-derived tracer images were obtained in neuromaps (https://github.com/netneurolab/neuromaps). Neuroimaging data of AD patients were obtained in ADNI 2 (https://adni.loni.usc.edu/data-samples/adni-data). Neuroimaging data of ASD patients were obtained in ABIDE I (https://fcon_1000.projects.nitrc.org/indi/abide/). Neuroimaging data of MDD patients were publicly available (https://www.synapse.org/Synapse:syn22317081). The NeuroSynth activation maps were publicly available at https://www.neurosynth.org. The source data underlying the figures can be found in Supplementary Data. Code availability All code used to perform the analyses is available at https://github.com/Tito11-creater/GM-WM-FC. Matlab (R2020a version, https://www.mathworks.com/products/matlab.html) and R (2023.06.1 + 524 version, https://dailies.rstudio.com/version/2023.06.1 + 524/) were used for running scripts. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s42003-026-10483-7. References Associated Data Supplementary Materials Supplementary Material Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Data Availability Statement Structural and functional neuroimaging data of healthy participants were obtained through the Washington University-University of Minnesota (WU-Minn) Human Connectome Project (https://www.humanconnectome.org/software/connectomedb). PET-derived tracer images were obtained in neuromaps (https://github.com/netneurolab/neuromaps). Neuroimaging data of AD patients were obtained in ADNI 2 (https://adni.loni.usc.edu/data-samples/adni-data). Neuroimaging data of ASD patients were obtained in ABIDE I (https://fcon_1000.projects.nitrc.org/indi/abide/). Neuroimaging data of MDD patients were publicly available (https://www.synapse.org/Synapse:syn22317081). The NeuroSynth activation maps were publicly available at https://www.neurosynth.org. The source data underlying the figures can be found in Supplementary Data. All code used to perform the analyses is available at https://github.com/Tito11-creater/GM-WM-FC. Matlab (R2020a version, https://www.mathworks.com/products/matlab.html) and R (2023.06.1 + 524 version, https://dailies.rstudio.com/version/2023.06.1 + 524/) were used for running scripts.