Introduction
The human brain is a complex, controllable system of interconnected neural networks across multiple spatiotemporal scales (Padamsey & Rochefort, Reference Padamsey and Rochefort2023; Seguin, Sporns, & Zalesky, Reference Seguin, Sporns and Zalesky2023). Dynamic shifts in brain state are essential for adaptive behavior, and these dynamic transitions are both supported and constrained by the structural architectures of neural networks (Betzel, Gu, Medaglia, Pasqualetti, & Bassett, Reference Betzel, Gu, Medaglia, Pasqualetti and Bassett2016; Kim et al., Reference Kim, Soffer, Kahn, Vettel, Pasqualetti and Bassett2017; Luppi et al., Reference Luppi, Singleton, Hansen, Jamison, Bzdok, Kuceyeski and Misic2024). Inadequate and abnormal state transitions linked to cognitive and behavioral symptoms of major depressive disorder (MDD) (J. Li et al., Reference Li, Wang, Xia, Zhang, Meng, Xu and Liao2024; Q. Li et al., Reference Li, Zhao, Hu, Liu, Wang, Zhang and Gong2024; Pan et al., Reference Pan, Ma, Wang, Zhang, Wang and Zhang2024). However, the neurobiological mechanisms by which network architectures shape dysfunctional dynamic transitions in MDD have not been extensively examined.
The network control theory (NCT) takes a systematic engineering approach and uses fundamental principles of cybernetics to provide insights into how the brain governs complex network dynamics (Medaglia, Pasqualetti, Hamilton, Thompson-Schill, & Bassett, Reference Medaglia, Pasqualetti, Hamilton, Thompson-Schill and Bassett2017; Srivastava, Fotiadis, Parkes, & Bassett, Reference Srivastava, Fotiadis, Parkes and Bassett2022). The concept of brain state is central to NCT, defined as the patterns of neural activity across brain regions or voxels at a given moment (Wu, Huang, Wang, & He, Reference Wu, Huang, Wang and He2024). The NCT leverages linear diffusion to model network communication, with external control directing transitions from initial to target brain states (D’Souza, di Bernardo, & Liu, Reference D’Souza, di Bernardo and Liu2023; Parkes et al., Reference Parkes, Kim, Stiso, Brynildsen, Cieslak, Covitz and Bassett2024). Therefore, the NCT provides a framework to characterize how the structural network architecture supports dynamic transitions across brain states (Deng, Li, Thomas Yeo, & Gu, Reference Deng, Li, Thomas Yeo and Gu2022; Gu et al., Reference Gu, Pasqualetti, Cieslak, Telesford, Yu, Kahn and Bassett2015; Xia et al., Reference Xia, Liu, Li, Meng, Yang, Chen and Liao2024). Two commonly used controllability metrics were average controllability and modal controllability. The average controllability of a network reflects the averaged input energy from a control node set and all possible states. Central nodes with high average controllability efficiently drive the brain to easily accessible states with minimal energy input. Cognitively, these nodes support multitasking and low-load processes by facilitating efficient state shifts. Modal controllability measures the ability of each node to control a specific, hard-to-reach state transition. Unlike central hubs, nodes with high modal controllability often exhibit moderate connectivity and facilitate transitions to cognitively demanding states (Gu et al., Reference Gu, Pasqualetti, Cieslak, Telesford, Yu, Kahn and Bassett2015; Karrer et al., Reference Karrer, Kim, Stiso, Kahn, Pasqualetti, Habel and Bassett2020; Wu et al., Reference Wu, Huang, Wang and He2024). Controllability is an energy-based measure showing heritable properties and biological relevance, such as oxygen and glucose metabolism and neurotransmitter receptor density (Ceballos et al., Reference Ceballos, Luppi, Castrillon, Saggar, Misic and Riedl2025; Frangou, Bassett, Glahn, Rodrigue, & Lee, Reference Frangou, Bassett, Glahn, Rodrigue and Lee2020; Gu et al., Reference Gu, Pasqualetti, Cieslak, Telesford, Yu, Kahn and Bassett2015; He et al., Reference He, Caciagli, Parkes, Stiso, Karrer, Kim and Bassett2022; Luppi et al., Reference Luppi, Singleton, Hansen, Jamison, Bzdok, Kuceyeski and Misic2024).
Numerous studies have used NCT to identify structural network controllability changes in neurological and psychiatric disorders, such as epilepsy (Bernhardt et al., Reference Bernhardt, Fadaie, Liu, Caldairou, Gu, Jefferies and Bernasconi2019; He et al., Reference He, Caciagli, Parkes, Stiso, Karrer, Kim and Bassett2022; Janson et al., Reference Janson, Sainburg, Akbarian, Johnson, Rogers, Chang and Morgan2024), schizophrenia (Li et al., Reference Li, Yao, You, Liu, Deng, Li and Gong2023; Wang et al., Reference Wang, Zhang, Yu, Niu, Niu, Li and Liu2022), bipolar disorder (Wang et al., Reference Wang, Zhang, Yu, Niu, Niu, Li and Liu2022), and associated these abnormalities with cognitive and behavioral function. Changes in network controllability correlate with core MDD clinical features and predict symptom improvement following treatment (Fang et al., Reference Fang, Godlewska, Cho, Savitz, Selvaraj and Zhang2022; Li et al., Reference Li, Wang, Xia, Zhang, Meng, Xu and Liao2024). In addition, regional network controllability associated with genetic, individual, and familial risk factors for MDD (Hahn et al., Reference Hahn, Winter, Ernsting, Gruber, Mauritz, Fisch and Repple2023). Collectively, these findings suggest that NCT can help elucidate the pathophysiological basis of cognitive and behavioral dysfunctions in MDD and aid in the formulation of effective individualized therapies (Liu, Slotine, & Barabási, Reference Liu, Slotine and Barabási2011; Pan et al., Reference Pan, Ma, Wang, Zhang, Wang and Zhang2024).
Previous studies on network controllability have constructed structural networks using diffusion tensor imaging (DTI) tractography (Parkes et al., Reference Parkes, Kim, Stiso, Brynildsen, Cieslak, Covitz and Bassett2024; Tang et al., Reference Tang, Giusti, Baum, Gu, Pollock, Kahn and Bassett2017). However, DTI tractography cannot provide direct metrics of tract connectivity strength (Donahue et al., Reference Donahue, Sotiropoulos, Jbabdi, Hernandez-Fernandez, Behrens, Dyrby and Glasser2016; Lewis et al., Reference Lewis, Li, Zhang, Meng, Yang, Xia and Liao2024). Importantly, the critical yet overlooked role of gray matter in controlling brain dynamics warrants emphasis, as it constitutes the primary energy-consuming substrate of the brain. Research has shown that network control ability depends on complementary contributions of white matter connectivity and gray matter volume, which together underlie the brain’s control properties (Jamalabadi et al., Reference Jamalabadi, Zuberer, Kumar, Li, Alizadeh, Amani and Walter2021). Recently, morphometric inverse divergence (MIND) was proposed to estimate within-subject similarities in gray matter morphometric parameters among regions for defining morphometric networks (Sebenius et al., Reference Sebenius, Seidlitz, Warrier, Bethlehem, Alexander-Bloch, Mallard and Morgan2023). MIND networks exhibit biological associations with tract-tracing measures of axonal connectivity, cortical cytoarchitectonics, gene expression patterns, and genetic heritability (Sebenius et al., Reference Sebenius, Seidlitz, Warrier, Bethlehem, Alexander-Bloch, Mallard and Morgan2023). Therefore, under the constraints of structural connection, exploring MIND network controllability would provide a more comprehensive understanding of how structural architectures shape and constrain the dynamic transitions in brain states.
In this study, we investigated the brain’s controllability of morphometric similarity networks in MDD and the underlying neurobiological mechanisms (see Figure 1 for analytic workflow). First, combining the structural and MIND networks, we constructed a morphometric network for each MDD patient and healthy controls (HCs). Second, we identified MDD-related abnormalities in the controllability of morphometric similarity networks using two independent cohorts. Third, we evaluated cognitive and biobehavioral associations of MDD-related network controllability utilizing the Neurosynth dataset. Finally, we examined spatial associations between changes in MDD-related morphometric network controllability and molecular signatures, including neurotransmitter pathways, metabolic distribution profiles, and regional gene transcription patterns.
Figure 1. Study outline. (A) Using diffusion and structural MRI data, DTI and MIND networks were constructed for each MDD patient and HC using the HCP_MMP atlas. The morphometric network was then constructed by merging the DTI and MIND networks. Regional average controllability and modal controllability were evaluated for the morphometric network using network control theory. (B) MDD-related alterations in regional controllability were then calculated as Cohen’s d values and mapped. Next, associations of regional controllability changes with cognitive and biobehavioral topics from the Neurosynth meta-analysis list were evaluated. Multiple linear regression analyses were performed to evaluate associations with neurotransmitomic and metabolic profiles. A PLS regression analysis was conducted to reveal regional gene expression patterns (from the AHBA dataset) associated with MDD-related changes in network controllability. Finally, gene set enrichment analyses were performed for GO terms, brain cell types, and chromosomes.
Methods
Participants
Individuals with MDD and age- and sex-matched HCs in the discovery cohort were recruited from the Southwest University Center for Brain Imaging. MDD patients and HCs in the replication cohort were acquired from the Department of Psychiatry, First Hospital of Shanxi Medical University, and the community nearby. Depression severity was assessed with the 17-item Hamilton Depression Rating Scale (HAMD). All MDD patients were diagnosed using the Structural Clinical Interview for Diagnostic and Statistical Manual of Mental Disorders-IV (DSM-IV) by experienced psychiatric physicians. Exclusion criteria for MDD patients were <18 years of age, major neurological or other psychiatric disorders, magnetic resonance imaging (MRI) abnormalities, and contraindications for MRI, such as metal or electronic implants. Inclusion criteria for HCs were (1) no axis I psychiatric disorders or neurological disorders, (2) no axis II antisocial or borderline personality disorders, and (3) no history of psychiatric illness among first-degree relatives.
This study was approved by the Ethics Committee of Southwest University and the First Affiliated Hospital of Chongqing Medical University, and the local ethics committee of the Shanxi Medical University. All study protocols were performed according to the Helsinki Declaration of 1975, and informed written consent was obtained from all participants before examinations.
MRI data acquisition and preprocessing
All MRI images in the discovery cohort were acquired using a Siemens Trio 3.0 T scanner at the Southwest University Center for Brain Imaging. The acquisition parameters for three-dimensional T1w images were as follows: repetition time (TR) = 1900 ms, echo time (TE) = 2.52 ms, field of view (FOV) = 256 × 256 mm2, matrix size = 256 × 256, flip angle = 9°, and voxel size =1 × 1 × 1 mm3. Diffusion MRI images were acquired using a diffusion-weighted, single shot, spin-echo, gradient-echo planar imaging sequence (TR = 11,000 ms, TE = 98 ms, FOV = 256 × 256 mm2, matrix size = 128 × 128, voxel size =2 × 2 × 2 mm3, slices = 60, one volume with b = 0 s/mm2, and 30 noncollinear directions b = 1000 s/mm2).
All images in the replication cohort were acquired with a 3.0-T Trio Siemens System at the Shanxi Provincial People’s Hospital, Taiyuan, China. Structural MRI images were acquired using a 3D magnetization-prepared rapid gradient-echo T1-weighted sequence: TR = 2300 ms, TE = 2.95 ms, TI = 900 ms, flip angle = 9°, FOV = 225 × 240 mm2, 160 slices, and thickness = 1.2 mm. For each DTI scan, 45 contiguous axial slices were acquired with the following parameters: TR = 6000 ms, TE = 90 ms, flip angle = 90°, slice thickness = 3 mm, FOV = 240 mm × 240 mm2, matrix size =128 × 128, and voxel size = 1.875 × 1.875 × 3 mm3. Diffusion-sensitizing gradients were applied along 12 noncollinear directions (b = 1000 s/mm2), together with a nondiffusion-weighted acquisition (b = 0 s/mm2).
The T1w images were preprocessed in surface-based space using FreeSurfer v6.0. Briefly, the cortical surface of each participant was reconstructed from 3D images by skull stripping, motion correction, classification of cortical gray matter and white matter, coordinate transformation, separation of hemispheres, subcortical structures, and subsequent construction of the gray–white interface and pial surface. The diffusion MRI images were preprocessed using the MRtrix3 package (Tournier et al., Reference Tournier, Smith, Raffelt, Tabbara, Dhollander, Pietsch and Connelly2019). The specific processes include Marchenko-Pastur Principal Component Analysis (MP-PCA) denoising, Gibbs ringing artefacts removing, eddy current, motion and susceptibility-induced distortion correction, bias field correction, registration with T1w images, T1 tissue segmentation, response function estimation, and multitissue constrained spherical deconvolution. Whole-brain probabilistic tractography was performed with 10 million streamlines, utilizing anatomically constrained tractography (ACT) framework. Finally, spherical-deconvolution informed filtering of tractograms (SIFT2) was performed to reduce and reconstruct streamlines weighted by cross-section multipliers (Smith, Tournier, Calamante, & Connelly, Reference Smith, Tournier, Calamante and Connelly2015).
Morphometric network controllability
An individual-specific structural connectivity network (SCN) was constructed by mapping reconstructed fiber pathways to the HCP_MMP atlas, and the connection strength was quantified as the number of streamlines. Consequently, the SCN was binarized to indicate the presence or absence of white matter fiber tract connections across cortical regions. Individual MIND network was constructed by evaluating morphological similarity across cortical regions (Supplementary Methods). Finally, the SCN thresholds the MIND network to construct a morphological network that simultaneously reflects the connections of white matter fiber bundles and the biological properties of gray matter.
NCT provides ways to assess the capacity of a given brain region to drive the entire system to a target state. To quantify regional controllability, a noiseless linear time-invariant brain structure controllability model was first constructed according to the relationship (Sun et al., Reference Sun, Jiang, Dai, Dufford, Noble, Spann and Scheinost2023):
$$ \mathrm{x}\left(t+1\right)=\mathrm{Ax}(t)+{B}_K(t){u}_K(t), $$
where x represents the brain state at time t, matrix A represents the morphological network, matrix BK describes the brain region (set to 1) injected with activity, induces state transition, and
$ {u}_K $
represents the input control strategy over time.
Two nodal controllability measures were calculated: average and modal controllability. Average controllability is estimated as the trace of the controllability Gramian matrix, Trace (WK). In classic control theory, WK is defined as:
$$ {W}_K={\sum}_{\tau =0}^{\infty }{A}^{\tau }{B}_K{B}_K^T{A}^{\tau }, $$
where T represents the transpose operation and τ represents the time step of the trajectory (Gu et al., Reference Gu, Pasqualetti, Cieslak, Telesford, Yu, Kahn and Bassett2015). Modal controllability is estimated as:
$$ {\varnothing}_i={\sum}_{j=1}^N\left(1-{\lambda}_j^2(A)\right){\unicode{x03BD}}_{ij}^2, $$
where λj and νij represent eigenvalues and elements of the eigenvector matrix of A, respectively. The average controllability and modal controllability of each node in A were calculated to construct individual participant models for group comparisons.
Case–control comparison analysis
To identify MDD-related abnormalities in brain controllability, semiparametric generalized additive models were used to estimate case–control differences in nodal- and network-level controllability:
$$ Y={\beta}_0+{\beta}_1\cdotp group+{f}_1(age)+{\beta}_2\cdotp gender+{\beta}_3\cdotp TIV, $$
where Y represents the measure of nodal- and/or network-level controllability. A smooth function f 1() was introduced as a non-parametric term to control for the effect of age, allowing flexible evaluation of nonlinear relationships without needing a prior shape. Cohen’s d values were calculated to represent the effect sizes of the group comparisons, and multiple comparisons were corrected by the false discovery rate (FDR) method (P < 0.05). For group comparisons, the 360 cortical regions were divided into 12 intrinsic functional networks and seven cytoarchitectonic networks (Triarhou, Reference Triarhou2007). The same model was used to assess MDD-related alterations in controllability among brain networks and von Economo cytoarchitectonic classes. The semiparametric generalized additive model was constructed with the mgcv R package v1.9 (https://cran.rproject.org/web/packages/mgcv/index.html).
Cognitive and biobehavioral association analyses
To assess the cognitive impact of regional MDD-related abnormalities in controllability, we quantified their spatial overlap with cognitive brain maps and conducted subsequent meta-analyses for each cognitive domain. Twenty-three cognitive terms were selected from the Neurosynth 50 meta-analytic topics (www.neurosynth.org/analyses/topics/v5-topics-50/) (Liu, Xia, Wang, Liao, & He, Reference Liu, Xia, Wang, Liao and He2020; Yarkoni, Poldrack, Nichols, Van Essen, & Wager, Reference Yarkoni, Poldrack, Nichols, Van Essen and Wager2011). To decode the biobehavioral correlates of whole-brain controllability in MDD, we spatially correlated its patterns with each of the 91 meta-analytic topics from the 200-topic Neurosynth meta-analysis list, derived via multilevel kernel density analysis of chi-square algorithm (Lotter et al., Reference Lotter, Kohl, Gerloff, Bell, Niephaus, Kruppa and Konrad2023). Empirical P values were derived by comparing these correlation coefficients to those derived from 10,000 null maps corrected for spatial autocorrelation and the FDR procedure (P < 0.05).
Correlation with neurotransmitter and metabolism distribution profiles
Multiple linear regression models were used to investigate the associations between MDD-related controllability abnormalities and neurotransmitter and metabolic distribution profiles (Supplementary Methods). The responder variable was the Cohen’s d value for each brain region, and the predictor variables included neurotransmitter receptors and metabolic profiles. After quantifying the explained variance, the relative contribution of each predictor variable (relative importance of regressor) was assessed using the pmvd metric in the relaimpo R package v2.2.7 (http://prof.beuth-hochschule.de/groemping/software/relaimpo/) (Gromping, Reference Gromping2006). The robustness and significance of these importance metrics were evaluated via bootstrapping, from which 95% confidence intervals were derived.
Correlation with gene expression profiles
Brain-wide gene expression data were extracted from the Allen Human Brain Atlas (AHBA) and preprocessed using the abagen toolbox to obtain a region-by-gene expression matrix (180 × 7411) for further analysis (Supplementary Methods) (Markello et al., Reference Markello, Arnatkeviciute, Poline, Fulcher, Fornito and Misic2021). A partial least squares (PLS) regression analysis was then constructed to identify weighted linear combinations of gene expression patterns associated with MDD-related controllability abnormalities. A permutation analysis corrected for spatial autocorrelation (https://github.com/frantisekvasa/rotate_parcellation) was used to test the statistical significance of the variance explained by PLS components. Subsequently, bootstrapping resampling (1000 times) was used to estimate the contribution of each gene. Z-scores were computed as each gene’s weight divided by its bootstrap standard error. Genes were then ranked based on their z-scores, and the significance was evaluated using one-sample Z tests. Finally, genes that remained significant after FDR correction (P < 0.05) were divided into PLS+ and PLS− groups.
Gene set enrichment analysis
To identify biological associations potentially linked to the PLS gene set (PLS+ and PLS−), functional enrichment analysis for gene ontology (GO) gene sets was performed with Metascape (Zhou et al., Reference Zhou, Zhou, Pache, Chang, Khodabakhshi, Tanaseichuk and Chanda2019). Enriched terms were clustered based on pairwise similarity (Kappa-test), and the resulting pathways were filtered at an FDR-corrected significance threshold of 0.05.
Based on the integrated single-cell transcriptomic and epigenomic data of adult brain, (Lake et al., Reference Lake, Chen, Sos, Fan, Kaeser, Yung and Zhang2017), we extracted cell-specific gene sets of 25 brain cell types, mainly including astrocytes, microglia, endothelial cells, oligodendrocyte precursors, oligodendrocytes, and neuronal cells. Enrichment scores for each set were normalized against a permutation-based null model (n = 10,000), yielding a normalized enrichment score (NES) corrected for gene set size. Similarly, GSEA analysis was used to calculate the NES for chromosomes 1:22 and X, Y (https://david.ncifcrf.gov/), and the FDR correction (P < 0.05) was performed. The GSEA was conducted with the clusterProfiler R package v4.8.3 (https://bioconductor.org/packages/release/bioc/html/clusterProfiler.html).
Sensitivity and validation analysis
Given that total intracranial volume (TIV) is an essential factor in brain volumetric analyses (Malone et al., Reference Malone, Leung, Clegg, Barnes, Whitwell, Ashburner and Ridgway2015), we first assessed its group-level difference between MDD patients and HCs, and its effect on controllability comparisons. To establish the robustness of the MIND network, we employed a leave-one-feature-out approach by iteratively excluding each MRI feature during network construction (Li et al., Reference Li, Keller, Seidlitz, Chen, Li, Weng and Liao2023; Sebenius et al., Reference Sebenius, Seidlitz, Warrier, Bethlehem, Alexander-Bloch, Mallard and Morgan2023; Yang et al., Reference Yang, Wagstyl, Meng, Zhao, Li, Zhong and Liao2021). Finally, given prior evidence of antidepressant effects on brain controllability (Fang et al., Reference Fang, Godlewska, Cho, Savitz, Selvaraj and Zhang2022; Piccinini et al., Reference Piccinini, Sanz Perl, Pallavicini, Deco, Kringelbach, Nutt and Tagliazucchi2025), we stratified MDD patients into medicated and unmedicated subgroups to specifically evaluate the impact of medication.
Furthermore, we assessed the consistency of case–control differences in network controllability between the discovery and an independent dataset. We reconstructed morphological networks for each subject in the replication cohort to compute regional average and modal controllability, derived case–control differences (Cohen’s d) using the same model (Equation 4). We evaluated spatial similarity between Cohen’s d maps. Spatial permutation testing (spin-test) was used to detect the significance of similarity after correction for multiple comparisons.
Results
After quality control and exclusion (Supplementary Methods), 173 patients with MDD and 191 HCs were included in the discovery cohort, and 146 MDD patients and 105 HCs were included in the replication cohort. There were no significant group differences in age and sex ratio (Supplementary Table S1). In both groups, we found a negative correlation between average controllability and modal controllability (Supplementary Figure S1), suggesting that these metrics are not independent in brain networks. Therefore, our primary analyses focused on average controllability, while findings for modal controllability are provided in Supplementary Materials (Supplementary Results; Supplementary Figures S2–S6).
MDD-related abnormalities in average controllability
The distribution patterns of average controllability in MDD and HC were spatially similar (R (358) = 0.99, P spin < 0.001), with higher average controllability in the posterior cingulate cortex, somatosensory and motor cortex, and visual regions, but lower average controllability in the medial temporal cortex, insula, and frontal opercular cortex (Figure 2A and B).
Figure 2. Differences in average controllability between patients with MDD and healthy controls. (A) Spatial distribution patterns of average controllability in healthy controls (HCs). (B) Spatial distribution patterns of average controllability in MDD patients. (C) MDD-related alterations of regional average controllability (versus HCs) expressed as a Cohen’s d map. Cortical regions showing statistically significant differences are circled (P FDR < 0.05). (D) MDD-related network differences in average controllability. (E) MDD-related alterations in average controllability for von Economo cytoarchitectonic classes. An asterisk represents significant differences (P < 0.05 after FDR correction).
Compared to HCs, MDD patients exhibited greater average controllability in the visual cortex, somatosensory cortex, and part of the inferior parietal cortex, but lower average controllability in the medial prefrontal and lateral temporal cortex (Figure 2C; Supplementary Table S2). At the global level, MDD patients exhibited significantly reduced whole-brain average controllability (Cohen’s d = −0.23, P = 0.03; Supplementary Figure S7A). At the network level, regions showing decreased average controllability in MDD were primarily located in the frontoparietal and default mode networks (Figure 2D). In von Economo classes, individuals with MDD demonstrated lower average controllability in association with the Economo cytoarchitectonic class, whereas individuals with MDD demonstrated higher average controllability in the primary sensory class (Figure 2E; Supplementary Table S4).
PLS regression identified greater average controllability of all cortical regions as an independent predictor of higher HAMD score (indicative of more severe symptoms) (R (171) = 0.38, P perm < 0.001). In addition, the visual cortex, posterior cingulate cortex, and part of the inferior parietal cortex showed higher weight for HAMD prediction (Supplementary Figure S8A and B).
Cognitive and biobehavioral terms associated with regions showing MDD-related abnormalities in average controllability
The regions showing increased average controllability in MDD were associated with ‘vision and sensory’, ‘action’, and ‘language’. Brain regions with reduced average controllability were associated with ‘memory’ and ‘social’ (Figure 3A). At the whole-brain level, we observed significant associations between the change in average controllability and topics related to ‘visual’, ‘sensory’, and ‘cognition’ domains (Figure 3B).
Figure 3. Functional decoding of average controllability differences using Neurosynth topics. (A) Bar charts of cognitive terms associated with regions showing significantly higher (left) and lower (right) average controllability in MDD. (B) Biobehavioral associations with regional MDD-related alterations in average controllability. Point sizes and colors represent nonparametric P-values of Spearman correlations between whole-brain differences and meta-analytic topic maps. Dashed lines represent P < 0.05. FDR-corrected significant (P < 0.05) topics are annotated (*).
Neurotransmitomic and metabolic signatures for MDD-related average controllability
The model of neurotransmitter receptor/transporter explained 5.85% of the variance in MDD-related average controllability (F (7, 352) = 3.12, P = 0.003). The predicted average controllability differences were positively correlated with observed differences (R (358) = 0.24, P spin = 0.0019, Figure 4A and B), underscoring the utility of the model, and the serotonin system showed the most substantial relative influence on observed differences (relative contribution = 41.0%, P FDR < 0.05; Figure 4C; Supplementary Table S6).
Figure 4. Neurotransmitomic and metabolic signatures of case–control differences in average controllability. (A) The predicted average controllability difference (Cohen’s d map) with neurotransmitter profiles. (B) The predicted d values with neurotransmitter profiles were positively correlated with the observed values. (C) The relative contribution of each neurotransmitter transporter/receptor to the multiple linear regression model. (D) The predicted average controllability difference (Cohen’s d map) with metabolic profiles. (E) The predicted d values based on metabolic profiles were positively correlated with the observed values. (F) The relative contribution of each metabolic system to the multiple linear regression model. An asterisk indicates significance after FDR correction (P < 0.05).
The model of brain metabolic patterns explained 6.92% of the variance in MDD-related average controllability (F (5, 354) = 5.26, P = 0.0001). Again, the predicted differences in controllability were positively correlated with observed differences (R (358) = 0.26, P spin = 0.0012, Figure 4D and E), and oxygen and glucose metabolism were the strongest contributors (relative contribution: CMRGlu, 28.7%; CMRO2, 28.9%; both P FDR < 0.05; Figure 4F; Supplementary Table S7).
Transcriptomic signatures for MDD-related average controllability
The first component (PLS1) explained 14.6% of the variance in MDD-related average controllability changes (P spin = 0.03). Notably, we found a positive correlation between observed differences and the expression pattern of PLS1 (R (178) = 0.38, P spin = 0.003, Figure 5A–C). We ranked and corrected all genes by their weights, yielding 896 overexpressed genes (PLS1+ genes) and 1274 underexpressed genes (PLS1− genes) associated with regional changes in average controllability (Figure 5D). Furthermore, these PLS1 genes were enriched for biological processes and components related to protein and mitochondria such as ‘protein catabolic process’, ‘protein localization to organelle’, ‘mitochondrial matrix’, ‘mitochondrial membrane’, and ‘ribonucleoside triphosphate phosphatase activity’ (Figure 5E). The GSEA of cell types and chromosomes indicates that PLS1 genes were primarily involved in excitatory and inhibitory neurons (Figure 5F) and revealed overrepresentation of DEGs on chromosomes 22 and 18 (P FDR < 0.05, Figure 5G).
Figure 5. Transcriptomic signatures of case–control differences in average controllability. (A) The MDD-related average controllability alterations in the left hemisphere. (B) The weighted gene expression profile of PLS1. (C) Scatterplot showing that regional d values were positively correlated with regional PLS1 scores. (D) Ranked PLS1 genes according to their weights and divided into PLS1+ and PLS1- subgroups (P < 0.05 after FDR correction). (E) Gene set enrichment analysis for GO terms. The size of the circle represents the number of PLS1 genes overlapping with each term, and the color represents significance (P < 0.05 after FDR correction). (F) Gene set enrichment analysis for brain cell types. The horizontal axis represents the normalized enrichment score. (G) Gene set enrichment analysis for chromosomes. The terms, cell types, and chromosomes in bold are significant after FDR correction (P < 0.05).
Stability and reproducibility of MDD-related changes in average controllability
MDD patients exhibited reduced whole-brain volume, and the spatial patterns of average controllability differences were highly consistent, regardless of whether TIV was regressed out (R (358) = 0.998, P spin < 0.001; Supplementary Figure S9A and B). Across all leave-one-feature-out iterations, the pattern of average controllability differences remained highly consistent with the original findings, confirming the robustness of the MIND network (Supplementary Figure S10). Moreover, MDD patients were divided into medicated (n = 85, 54 females) and unmedicated (n = 80, 52 females) subgroups to compare their average controllability with that of HCs. The original pattern of case–control differences in average controllability remained robust and was independent of medication status (Supplementary Figure S11).
The spatial distributions of average controllability in both MDD patients and HCs were replicated across cohorts, demonstrating consistent elevation in sensorimotor, posterior cingulate, and visual cortices, alongside reduced controllability in medial temporal, insular, and frontal opercular regions (Supplementary Figure S13A and B). The identically derived case–control differences (d-map) from the replication cohort were spatially concordant with the discovery cohort (R (358) = 0.45, P spin < 0.001; Supplementary Figure S13C).
Discussion
We identified MDD-related regional abnormalities in morphometric network controllability and elucidated the underlying cellular and molecular mechanisms. These alterations spatially correlated with dopamine receptor/transporter densities, oxygen/glucose metabolism levels, and relevant gene expression – notably for genes on chromosomes 12 and X related to protein transport, signaling, and glial/neuronal function. These findings reveal potential neurocellular and molecular mechanisms underlying MDD-associated abnormalities in network controllability and brain-state transitions.
Network controllability is jointly determined by white matter connections and nodal gray matter volume sufficiency, reflecting the dual constraints of wiring and computing resources (Jamalabadi et al., Reference Jamalabadi, Zuberer, Kumar, Li, Alizadeh, Amani and Walter2021). Therefore, relying solely on the white matter network overlooks the role of gray matter in control dynamics and the functional significance of white matter microstructure. White matter enables impulse conduction across gray matter, holding valuable functional significance in normal brain operation and disease processes (Ji et al., Reference Ji, Cui, D’Arcy, Liao, Biswal, Zhang and Wang2025; Ji, Liao, Chen, Zhang, & Wang, Reference Ji, Liao, Chen, Zhang and Wang2017; Ji et al., Reference Ji, Ren, Li, Sun, Liu, Gao and Wang2018). Evidence has demonstrated that MDD involves simultaneous disruptions in the structure and function of both white and gray matter (Chai et al., Reference Chai, Sheline, Oathes, Balderston, Rao and Yu2023; Ji et al., Reference Ji, Sun, Hua, Zhang, Zhang, Bai and Wang2023; Zhuo et al., Reference Zhuo, Li, Lin, Jiang, Xu, Tian and Song2019). To systematically evaluate brain control properties, we integrated multimodal gray and white matter microstructure to elucidate the pathology of dynamic abnormalities in MDD.
MDD-related control abnormalities were most frequently detected in the frontoparietal and visual areas, which support vision, attention, and memory. Abnormal controllability suggests dysfunction in the white matter structural connections and the gray matter processing unit. Previous studies on MDD have revealed structural connections and functional abnormalities in these regions (Gong & He, Reference Gong and He2015; Kaiser, Andrews-Hanna, Wager, & Pizzagalli, Reference Kaiser, Andrews-Hanna, Wager and Pizzagalli2015; Mulders, van Eijndhoven, Schene, Beckmann, & Tendolkar, Reference Mulders, van Eijndhoven, Schene, Beckmann and Tendolkar2015). MDD patients tend to low-effort neural states, which may underlie their difficulties with concentration, loss of interest, and impaired task-switching (Fang et al., Reference Fang, Godlewska, Cho, Savitz, Selvaraj and Zhang2022; Haroz et al., Reference Haroz, Ritchey, Bass, Kohrt, Augustinavicius, Michalopoulos and Bolton2017; Schultz et al., Reference Schultz, Ito, Solomyak, Chen, Mill, Anticevic and Cole2019). Correspondingly, reduced controllability in temporal and prefrontal cortices has been linked to cognitive deficits and depressive symptoms (Fang, Gao, Schulz, Selvaraj, & Zhang, Reference Fang, Gao, Schulz, Selvaraj and Zhang2021). Collectively, the controllability anomalies in MDD stem from an impairment in information processing and transmission at key nodes. These impede effective transitions between neural states, ultimately resulting in behavioral and cognitive deficits.
Network controllability is strongly energy-dependent and is associated with neurotransmitter systems implicated in neuropsychiatric disorders such as MDD (Castrillon et al., Reference Castrillon, Epp, Bose, Fraticelli, Hechler, Belenya and Riedl2023; Hansen et al., Reference Hansen, Shafiei, Markello, Smart, Cox, Norgaard and Misic2022; Nimgampalle et al., Reference Nimgampalle, Chakravarthy, Sharma, Shree, Bhat, Pradeepkiran and Devanathan2023). Leveraging biologically verified MIND networks, controllability models, and neurotransmitter mapping, we identified specific associations between network control and the serotonin system. Selective serotonin reuptake inhibitors, commonly used antidepressants that regulate motivation and depressive symptoms (Mao, Fan, Feng, & Dai, Reference Mao, Fan, Feng and Dai2025), are consistent with the monoamine hypothesis that mutations in serotonin transporter and/or receptor genes contribute to MDD pathophysiology (Malhi & Mann, Reference Malhi and Mann2018; Marx et al., Reference Marx, Penninx, Solmi, Furukawa, Firth, Carvalho and Berk2023). Similarly, MDD-related controllability abnormalities were associated with altered glucose and oxygen metabolism. This aligns with the high metabolic demands of neural dynamics, in which glucose fuels oxidative metabolism to support ATP production, neurotransmitter synthesis, and neuroplasticity (Ceballos et al., Reference Ceballos, Luppi, Castrillon, Saggar, Misic and Riedl2025; He et al., Reference He, Caciagli, Parkes, Stiso, Karrer, Kim and Bassett2022; Jansen et al., Reference Jansen, Milaneschi, Schranner, Kastenmuller, Arnold, Han and Penninx2024). These results reveal physiologically grounded signatures of controllability deficits in MDD, advancing our understanding of its pathogenesis.
These MDD-linked alterations in controllability were strongly correlated with dysregulated expression of protein- and mitochondria-related genes involved in synaptic, neurodevelopmental, and neurotransmission processes (Howard et al., Reference Howard, Adams, Clarke, Hafferty, Gibson, Shirali and McIntosh2019; Wray et al., Reference Wray, Ripke, Mattheisen, Trzaskowski, Byrne, Abdellaoui and Sullivan2018). Pathologically, protein catabolism disrupts glutamate metabolism via ammonia production (Griffin, Liu, Bradshaw, & Wang, Reference Griffin, Liu, Bradshaw and Wang2018; Henter, Park, & Zarate, Reference Henter, Park and Zarate2021), while mitochondrial dysfunction impairs energy supply for neuronal signaling and exacerbates oxidative stress under psychological stress, contributing to depression-related structural, functional, and cellular brain alterations (Daniels, Olsen, & Tyrka, Reference Daniels, Olsen and Tyrka2020; Lin, Liu, Qin, & Wang, Reference Lin, Liu, Qin and Wang2022; Picard & Shirihai, Reference Picard and Shirihai2022). Diverse brain cell types have been implicated in the etiology of MDD (Nagy et al., Reference Nagy, Maitra, Tanti, Suderman, Théroux, Davoli and Turecki2020). Our results showed gene-set enrichment in neuronal cells, consistent with reported interneuron polygenic risk and neuronal differentially expressed genes (DEGs) (Anderson et al., Reference Anderson, Collins, Kong, Fang, Li, He and Holmes2020). Single-nucleus transcriptomics further revealed DEGs in excitatory and inhibitory neurons (Chatzinakos et al., Reference Chatzinakos, Pernia, Morrison, Iatrou, McCullough, Schuler and Daskalakis2023). Additionally, DEG enrichment on chromosome 22 aligns with pharmacogenetic studies linking this locus to escitalopram metabolism in MDD (Ji et al., Reference Ji, Schaid, Desta, Kubo, Batzler, Snyder and Weinshilboum2014).
Several limitations should be considered. Consistent with previous NCT studies, we constructed a noise-free linear time-invariant brain controllability model (Luppi et al., Reference Luppi, Singleton, Hansen, Jamison, Bzdok, Kuceyeski and Misic2024; Parkes et al., Reference Parkes, Kim, Stiso, Brynildsen, Cieslak, Covitz and Bassett2024; Sun et al., Reference Sun, Jiang, Dai, Dufford, Noble, Spann and Scheinost2023). Although brain dynamics are inherently nonlinear, these provide reasonable approximations and can outperform specific nonlinear models in capturing macroscopic brain activity (Ju & Bassett, Reference Ju and Bassett2020; Kringelbach & Deco, Reference Kringelbach and Deco2020; Nozari et al., Reference Nozari, Bertolero, Stiso, Caciagli, Cornblath, He and Bassett2023). In addition, a controllable system under linearization often remains locally controllable within its nonlinear regime (Zañudo, Yang, & Albert, Reference Zañudo, Yang and Albert2017). Developing appropriate nonlinear models will help promote a more comprehensive understanding of brain controllability (Wu et al., Reference Wu, Huang, Wang and He2024). Second, our results are based on group-level comparisons of network controllability and neurobiological signatures. Further research is warranted to investigate the nuanced relationship between controllability and brain neurotransmitter and metabolic patterns at the individual level. Finally, therapeutic intervention in psychiatry can be conceptualized as an attempt to control dynamic transitions in brain network states (Braun et al., Reference Braun, Schaefer, Betzel, Tost, Meyer-Lindenberg and Bassett2018; Lynn & Bassett, Reference Lynn and Bassett2019). Indeed, recent studies have concluded that differences in treatment response can be explained by disease-related differences in network controllability (Parkes et al., Reference Parkes, Moore, Calkins, Cieslak, Roalf, Wolf and Bassett2021). Therefore, future research should investigate whether interventions guided by NCT are more effective than current approaches. Additionally, new tools are required to identify regions and state trajectories that must be controlled for effective targeted therapy.
Conclusion
This study investigated morphometric network controllability abnormalities related to MDD. These abnormalities were related to sensory perception and memory domains, and spatially associated with alterations in gene expression, neurotransmitter signaling, and metabolism, supporting the idea that multiscale organizations are related to neuroimaging-derived features of MDD, such as neuronal cells, serotonin, and oxygen and glucose metabolism. More broadly, our results highlight the advantages of using network control model-based approaches to understand the psychopathology of MDD.
Supplementary material
The supplementary material for this article can be http://doi.org/10.1017/S0033291725103140.
Acknowledgements
The authors are grateful to all the participants in this study. The authors thank International Science Editing (http://www.internationalscienceediting.com) for editing this manuscript.
Author contribution
Jinpeng Niu, Huafu Chen, Jiao Li, and Wei Liao have contributed to literature review, study design, and manuscript preparation. Yaohui He, Wei Li, Kangjia Chen, Qingjin Liu, and Jiang Qiu have contributed to study design, manuscript writing, data cleaning, and analyses. Jinpeng Niu, Jie Xia, and Wenxia Li provided inputs on study design and manuscript writing. All authors contributed to and have approved the final manuscript.
Funding statement
This work was supported by the National Natural Science Foundation of China (62571105, 62473082, 62036003, and 82121003) and the Fundamental Research Funds for the Central Universities (grant no. ZYGX2022YGRH008, and ZYGX2024XJ054).
Competing interests
The authors declare none.
Open access
