Introduction

Major depressive disorder (MDD) is among the most common and burdensome psychiatric disorders globally [1]. In addition to the clinical symptoms, including low mood, loss of interest, and fatigue, neuropsychological studies suggest that patients with MDD exhibit deficits in low-level sensory processing, such as visual processing and pain perception [2,3,4], and higher-order cognitive functions, such as memory, social interaction, and reasoning [5,6,7,8]. Although many prior studies reported widespread abnormalities and treatment effects in the brain structure and function in MDD [9,10,11,12,13,14], the neurobiological mechanism underlying these sensory-cognitive deficits remains to be elucidated.

A hierarchical architecture is one of the fundamental organizational principles of the human brain, allowing information encoding and integration from sensation to cognition [15]. Using resting-state functional MRI (R-fMRI) [16] and the gradient decomposition framework [17], researchers can decompose the functional brain networks into different gradient components to capture the topography of the connectomes. Such a gradient-based method presents the relative differences in connectivity patterns among regions along a continuous spectrum [17, 18]. In healthy adults, the principal gradient spans the primary sensory networks to the transmodal default mode network (DMN), which corresponds to the hierarchical synaptic distributions in postmortem brains [15]. This spatial pattern also captures the functional spectrum of activation maps from perception to abstract cognition [17] and is spatially similar to the cortical myelin content map [19]. Although many R-fMRI studies revealed alterations in functional activity and connectivity involving the primary [9, 20,21,22] and transmodal [20, 21, 23,24,25] systems in patients with MDD, no studies reported whether and how the principal primary-to-transmodal gradient of the functional connectome is disrupted in this clinical population and whether this disruption provides valuable information for treatment outcomes. The characterization of the principal connectome gradient in MDD could provide insight into the hierarchical network mechanisms for understanding the concurrence of sensory-cognitive impairments in MDD patients and offer potential biomarkers for treatment evaluations in this disorder.

Many studies indicated that MDD is a moderately heritable disorder [26]. Genome-wide association studies identified several risk variants of genes linked to MDD, some of which play key roles in the biological functions of presynaptic differentiation and neuroinflammation [27]. Recently, the transcriptome-connectome association studies provide a critical opportunity to bridge the gap between the microlevel transcriptome profile and the macroscale brain network [28,29,30,31]. Specifically, the functional architectures of the connectomes are associated with gene expression profiles involving ion channel activity and oxidative metabolism [29, 32]. Therefore, if patients with MDD exhibit disturbances in the macroscale connectome gradient, these abnormalities might be associated with transcriptome profiles. The elucidation of such an association would enhance our understanding of the molecular genetic underpinnings of the dysfunctional connectome gradient in MDD.

To address these gaps, we employed a large multisite R-fMRI dataset of 2227 individuals and postmortem gene expression data to investigate the functional connectome gradients in MDD and establish their associations with the transcriptome profile and treatment outcomes. Specifically, we hypothesized that (i) the principal primary-to-transmodal gradient is disrupted in patients with MDD in both the global gradient topography and the focal gradient scores of primary and transmodal systems; (ii) the regions with MDD-related alterations of connectome gradient are associated with meta-analytic maps involving multiple functional domains, and with gene expression profiles enriched in particular biological processes (e.g., synapse-related functions); and (iii) the connectome gradients in MDD patients could predict clinical outcomes after antidepressant treatment.

Materials and methods

Imaging dataset and preprocessing

We used two imaging datasets with a strict quality control. Dataset 1 included 1148 patients with MDD and 1079 controls (Table 1) recruited from nine research centers through the Disease Imaging Data Archiving—Major Depressive Disorder Working Group (DIDA-MDD). Dataset 2 included 20 first-episode drug-naïve patients (Supplementary Table 1) recruited from Peking University Sixth Hospital [13]. Each patient in dataset 2 received an 8-week antidepressant treatment with escitalopram (an antidepressant of selective serotonin reuptake inhibitor, SSRI), and the treatment outcomes were recorded. All participants were diagnosed by experienced psychiatrists using structured clinical interviews. The patients met the Diagnostic and Statistical Manual of Mental Disorders-IV (DSM-IV) diagnostic criteria for MDD [33] and did not meet the criteria for any other Axis I disorder or current substance use. The severity of depression was rated using the Hamilton depression rating scale (HDRS) [34]. The controls did not have a current or lifetime history of Axis I disorder. The study was approved by the ethics committees of each research center, and written informed consent was obtained from each participant. For each participant, the R-fMRI data were obtained using 3.0-T MRI scanners (Supplementary Table 2) and were preprocessed using a standard pipeline described in our previous work [9].

Table 1 Demographic and clinical characteristics of the participants.
Full size table

Gene expression dataset and preprocessing

The microarray-based gene expression data were downloaded from the Allen Human Brain Atlas (AHBA) [28, 35]. The tissue samples in this dataset were collected from the brains of six adult donors (mean age: 42.5 years, 1 female, Supplementary Table 3). Given the large transcriptional differences between cortical and subcortical regions [35], we conducted gene expression analysis within the cortical regions. We preprocessed the gene expression microarray data according to a recommended pipeline, including mapping the samples to a cortical parcellation with 360 regions [36], probe reannotation and selection, and normalization across donors [37].

Connectome gradient analysis

For each individual, we first constructed the voxelwise functional network (18,933 nodes) and then applied the diffusion map embedding approach [17, 38] to estimate the connectome gradient. Briefly, the top 10% connections of each node were retained, and the cosine similarity between each pair of nodes was calculated. The similarity matrix was further scaled into a normalized angle matrix to avoid negative values [39, 40]. Then, diffusion map embedding was applied to capture the gradient components explaining the variance in the connectivity pattern of the functional connectome. The resultant gradient maps were further aligned across individuals using iterative Procrustes rotation [38]. Given that the principal primary-to-transmodal gradient is closely associated with the neuronal microstructure and cognitive functions [17, 18], we primarily focused on MDD-related alterations in the principal gradients. Three global metrics, including gradient explanation ratio, range, and variation, were calculated. Finally, we utilized ComBat to correct the site effects on the gradient maps and metrics [9, 41]. The case-control differences in the connectome gradient were assessed by using a general linear model (GLM) with age and sex as covariates. To further explore the influence of sex on the group gradient differences, we used an additional GLM to examine possible sex-by-group interactions. For the global gradient metrics, the statistical significance threshold was set to P < 0.05. For the regional gradient score maps, the significance threshold was set to P < 0.001 at the voxel level, followed by Gaussian random field (GRF) correction at the cluster level of P < 0.05 [42].

Association between the functional network topology and connectome gradients in MDD

We further examined the association between MDD-related alterations in the connectome gradients and the network topology changes. Specifically, we computed the clustering coefficient (Cp) and the characteristic shortest path length (Lp) of each participant in the abovementioned whole-brain networks. Notably, Cp and Lp quantify brain network segregation and integration, respectively [43, 44]. The between-group differences in Cp and Lp were assessed using a GLM after controlling for age and sex. For the network metrics exhibiting significant between-group differences, we further examined the across-subject correlations to the global gradient metrics.

Association analysis between meta-analytic cognitive terms and MDD-related gradient alterations

We used Neurosynth (https://neurosynth.org/) [45] to examine the association between the meta-analytic cognitive terms and the MDD-related gradient alterations. The thresholded Z-maps derived from the between-group comparisons of regional gradient scores were first divided into MDD-positive (i.e., MDD > controls) and MDD-negative (i.e., MDD < controls) maps. We used the “decoder” function in Neurosynth to examine the spatial correlations between these MDD-positive or MDD-negative maps and the meta-analytic map of each term in the database. The significance of the correlation coefficients was determined by permutation tests in which spatial autocorrelations were corrected [46].

Association analysis between gene expression and MDD-related gradient alterations

We used a partial least squares (PLS) regression analysis [47] to explore the association between the transcriptional profiles and between-group differences in the principal gradient map. We first aligned the gene expression data (10,027 genes) and between-group difference Z-map of the principal gradient to a cortical parcellation atlas [36]. The gene expression data and the Z-map of the between-group differences of the principal gradient were set as the predictor variables and the response variable, respectively. We adopted a spatial autocorrelation corrected permutation test to examine whether the R2 of the PLS component was significantly greater than that expected by chance. For each significant component, we used a bootstrapping method to correct estimation error of the weight of each gene [48]. We then ranked the genes according to their corrected weights, which represent their contribution to the PLS regression component. Both the descending and ascending sequences were included in the gene enrichment analysis to identify enriched Gene Ontology terms by using GOrilla (http://cbl-gorilla.cs.technion.ac.il/) [49]. All three ontology categories, including biological process, molecular function, and cellular component, were considered. Significant enrichment was set to Benjamini-Hochberg false discovery rate (FDR)-corrected q < 0.05 [48, 50]. We also validated the sex and hemisphere effects on this connectome-transcriptome association analysis.

To further determine whether the MDD-related genes identified in previous gene studies contributed to the PLS model more than other disorder-related genes, we analyzed the disorder-related genes from the in situ hybridization (ISH) gene expression data provided by the AHBA (https://help.brain-map.org/display/humanbrain/Documentation). We selected previously reported MDD-related genes and genes related to other six brain disorders from the gene list. We compared the PLS weights between genes of MDD and those of other disorders using permutation tests. Moreover, we examined the correlation between the expression profile of each gene and the between-group Z-map of the principal gradient. The percentage of significantly correlated genes was compared between MDD and other disorders using permutation tests.

Effects of clinical factors

To investigate the effects of clinical factors on the connectome gradient, we classified the patients into different pairs of subgroups according to their clinical information, including patients with an early-onset age less than or equal to 21 years vs. greater than 21 years (this cutoff is consistent with the DSM-IV text revision definition of the early-onset of unipolar depressive disorders and previous studies [11, 51]; a cutoff age of 18 years was applied for validation), patients suffering from their first episode vs. recurrent episodes, and patients receiving medication vs. not receiving medication. Then, we compared the gradient metrics between each corresponding pair of subgroups by using a GLM. Moreover, we explored the relationships between the gradient metrics and HDRS using partial correlations with age and sex controlled.

Prediction of treatment outcome

A support vector regression (SVR) was conducted to examine the prediction ability of connectome gradient for treatment response (i.e., changes of HDRS) in patients. The baseline principle gradient maps were served as predictive features, and the model was validated using an embedded 5-fold cross-validation procedure and permutation tests. For comparison purposes, we also estimated the prediction ability of the functional connectivity strength (FCS, i.e., summed connectivity for each voxel) map for treatment response.

Validation analysis

We validated our results by considering six potential influencing factors, including specific scanning site, age, Fisher’s r-to-z transformation to the correlation matrix, head motion, different gradient alignment methods, and non-parametric permutation tests. The Jaccard index was used to examine the similarity between each validation result and the main finding.

Results

The principal primary-to-transmodal gradient explained 11.9 ± 3.1% of the total connectivity variance (MDD, 11.7 ± 3.2%; HC, 12.1 ± 3.0%, Supplementary Fig. 1), which was organized along a gradual axis from the primary visual/sensorimotor networks (VIS/SMN) to the DMN (Fig. 1a), replicating the recent observation of connectome gradients from the primary to the transmodal cortices in healthy adults [17]. The spatial patterns of the group-averaged principal gradient maps were remarkably similar between the MDD and HC groups (Spearman’s ρ = 0.999, P < 0.0001, permutation tests with spatial autocorrelation corrected, Supplementary Fig. 2). A visual inspection of the histogram revealed that the extremes of the primary-to-transmodal gradient were contracted in MDD relative to the control range (Fig. 1b). Specifically, for each system, we compared the gradient score between the group-averaged maps of the MDD group and control groups by using paired t-tests across the voxel. Compared with the HC group, the MDD group had higher gradient scores in the VIS and SMN but lower in the other subnetworks (FDR-corrected q < 0.05, Fig. 1c and Supplementary Table 4).

Fig. 1: Connectome gradient mapping in patients with MDD and controls.
figure 1

a The principal gradient was organized along a gradual axis from the primary visual/sensorimotor networks to the default mode network. Surface rendering was generated using BrainNet Viewer (www.nitrc.org/projects/bnv/) [68] with the inflated cortical 32 K surface [36]. b Global and c system-based histograms showing that the extreme values were contracted in the patients with MDD relative to those in the healthy controls. Arrows in c indicate the direction of the significant differences between the MDD patients and controls. The whole-brain voxels were assigned to eight systems according to a 7-system cortical parcellation [69] and a subcortical system of the AAL atlas [70]. VIS visual network, SMN sensorimotor network, DAN dorsal attention network, VAN ventral attention network, SUB subcortical regions, LIB limbic network, FPN frontoparietal network, DMN default mode network.

Full size image

Alterations in connectome gradients in MDD

The between-group statistical comparisons showed that the primary-to-transmodal gradient in the MDD group explained less variance in the functional connectome than that in the HC group (Cohen’s d = −0.16, P = 0.0002, Fig. 2a and Supplementary Table 5). Moreover, the patients with MDD showed a narrower range of scores (d = −0.21, P = 0.0000008) and less spatial variation (d = −0.20, P = 0.000002) than the HCs (Fig. 2a and Supplementary Table 5). Regionally, the MDD group showed lower gradient scores mostly in the DMN (76.6%) but higher scores mainly in the VIS (59.2%) and SMN (40.1%) than the HC group (absolute d = 0.18–0.25, voxel-level P < 0.001, GRF-corrected P < 0.05, Fig. 2b, Supplementary Table 6). No significant sex-group interaction effect was observed on either the global or local gradient measures. The results of the second and third gradients are shown in Supplementary Figs. 3, 4 and Supplementary Tables 5, 7, 8.

Fig. 2: Statistical comparison of the gradient metrics.
figure 2

a Case-control differences in the global metrics of the principal gradient. b Voxelwise statistical comparisons between the healthy controls and patients with MDD and distribution of the regional case-control difference in different systems. Higher/lower values in MDD are presented as warm/cold colors. The statistical significance level was set as voxel-level P < 0.001 and Gaussian random field cluster level-corrected P < 0.05. c Case-control differences in the clustering coefficient (Cp) and correlations between the Cp and global metrics of the principal gradient across all patients. d Word clouds of cognitive functions associated with brain regions that exhibited higher (red) or lower (blue) gradient scores in MDD. The font size of a given cognitive term corresponds to the correlation coefficient (r) of the between-group Z-map of the principal gradient and meta-analytic map of that term generated by Neurosynth. The boxes below the word clouds indicate the correspondence between the font size and the correlation coefficient. All correlations are significant (FDR q < 0.05). ***P < 0.001. VIS visual network, SMN sensorimotor network, DAN dorsal attention network, VAN ventral attention network, SUB subcortical regions, LIB limbic network, FPN frontoparietal network, DMN default mode network.

Full size image

Correlation between connectome gradients and network topology in MDD

The patients with MDD exhibited lower Cp than the HCs (d = −0.14, P = 0.0008), indicating lower functional segregation of brain networks in MDD. Moreover, Cp was correlated with the global metrics of the principal gradient across the individuals with MDD (explained ratio: r = 0.336, P < 0.001; gradient range: r = 0.257, P < 0.001; spatial variance: r = 0.356, P < 0.001, Fig. 2c).

Meta-analytic cognitive functions related to gradient alterations in MDD

The regions with higher principal gradient scores in MDD were correlated with several meta-analytic cognitive terms mainly involved in sensorial and perceptional processes, such as visual, sensory, and audiovisual processes (Fig. 2d). The regions with lower principal gradient scores in MDD were correlated with several high-order cognitive terms, including memory, social interaction, and reasoning (Fig. 2d). For details, see Supplementary Tables 9–11.

Gene expression profiles related to gradient alterations in MDD

The first two components of the PLS regression explained 51.7% of the variance in the MDD-related alterations in the principal gradient (P < 0.0001 for component 1 and P = 0.0037 for component 2, permutation tests with spatial autocorrelation corrected, Supplementary Fig. 5). Component 1 represented a transcriptional profile characterized by high expression mainly in the posterior parietal-occipital areas but low expression in the prefrontal areas (Fig. 3a). Component 2 revealed a gene expression profile with high expression mainly in the sensorimotor, visual, and temporal cortices but low expression in the frontoparietal cortices (Fig. 3b). The regional mapping of these two components was positively correlated with the Z-map of the primary-to-transmodal gradient (component 1: r = 0.574, P < 0.0001; component 2: r = 0.456, P < 0.0001, permutation tests with spatial autocorrelation corrected, Fig. 3a, b). The Gene Ontology enrichment analysis revealed that the genes ranked in ascending order of the component 1 weight were enriched in biological processes related to transsynaptic signaling and molecular function of calcium ion binding (FDR-corrected q < 0.05, Fig. 3c and Supplementary Table 12). No significant enrichment of cellular components was observed. Most terms remained significant in separate analyses of the male participants or left hemisphere (Supplementary Tables 13, 14). The genes ranked according to the weight of component 2 did not show significant enrichment. Moreover, the MDD-related genes from the ISH survey exhibited stronger PLS weights than the genes related to other disorders (except for autism, all Ps < 0.05, Fig. 3d). The expression profiles of sixty-seven percent (8 of 12) of the MDD-related genes were correlated with the between-group Z-map of the primary-to-transmodal gradient (Supplementary Table 15). This percentage was higher than that in all other disorder categories (all Ps < 0.0004, Fig. 3e).

Fig. 3: Association between MDD-related gradient alterations and gene expression.
figure 3

a The first PLS component (PLS 1) identified a gene expression profile with high expression mainly in the posterior parietal-occipital areas but low expression in the prefrontal areas. The transcriptional profiles were positively correlated with the between-group Z-map of the principal gradient (permutation tests with spatial autocorrelation corrected, 10,000 times). The shadow indicates the 95% confidence intervals. Each dot represents a region. b The second PLS component (PLS 2) revealed a gene expression profile with high expression mainly in the sensorimotor, visual, and temporal cortices but low expression in the frontoparietal cortices. The transcriptional profiles were positively correlated with the between-group Z-map of the principal gradient. c Genes ranked in ascending order of the PLS 1 weight were enriched in the biological process of transsynaptic signaling and molecular function of calcium ion binding (FDR q < 0.05). d The PLS weight of predefined genes related to different disorders based on the in situ hybridization (ISH) gene list provided by the AHBA. The asterisk indicates a significant difference between MDD and a specific disorder. **P < 0.01; *P < 0.05. e The percentage of disorder-related genes showing significant correlations between the expression profile and the between-group Z-map of the principal gradient. AD Alzheimer’s disease, ASD autism, EP epilepsy, ID intellectual disability, PD Parkinson’s disease, SCZ schizophrenia.

Full size image

Clinical relations to connectome gradients in MDD

The patients with an onset in adolescence (age ≤ 21 years, N = 300) showed a narrower gradient range (d = 0.29, P = 0.0006) and smaller region variation (d = 0.19, P = 0.018) in the principal primary-to-transmodal gradient than the patients with an onset age older than 21 years (N = 292) (Fig. 4a). These results were also replicated using a cutoff age of 18 years (Supplementary Table 16). There were no significant differences in the topographic features of the principal gradient between the patients who were and those who were not taking medication or the patients in their first episode and recurrent patients (Supplementary Table 16). The voxelwise comparisons showed that there was no significant difference in the regional scores of the principal gradient between these clinical category pairs. The range of the primary gradient was correlated with HDRS across all patients (r = 0.080, P = 0.009, Fig. 4b).

Fig. 4: Clinical effects on the gradient topography and prediction of treatment outcomes in patients with MDD.
figure 4

a Group differences in the range and variance of the principal gradient between the patients with an onset age ≤ 21 years and those with an onset age > 21 years. ***P < 0.001; *P < 0.05. b The correlation between the range of the principal gradient and the HDRS across all patients. c Scatter plot presents the correlation between the observed HDRS change after 8-week antidepressant treatment and the predicted HDRS change derived from the SVR analysis (permutation-tests; P < 0.001). Each dot represents the data from one patient, and the dashes indicate the 95% prediction error bounds. d The absolute summed weights in 5-fold cross-validation were mapped onto the brain surface. Regions with higher/lower predictive power were colored in white/red. The radar map represents the distribution of predictive power in different systems. OA onset age, HDRS Hamilton depression rating scale, VIS visual network, SMN sensorimotor network, DAN dorsal attention network, VAN ventral attention network, SUB subcortical regions, LIB limbic network, FPN frontoparietal network, DMN default mode network.

Full size image

Treatment outcome prediction of connectome gradients in MDD

The principal gradient map of the patients at baseline could significantly predict their HDRS changes after 8-week antidepressant treatment (r = 0.652, P < 0.001, one-tailed permutation test, Fig. 4c). The most contributive features were located in the DMN (percentage of total feature weights in SVR: 26.4%), SMN (15.3%), and FPN (15.2%) (Fig. 4d). In contrast, the baseline FCS map of the patients could not predict their HDRS changes (r = −0.385, P = 0.902, one-tailed permutation test).

Validation results

Overall, the MDD-related alterations in the principal connectome gradient observed using different validation strategies remained highly similar to our main findings, with a Jaccard index = 0.90 ± 0.12 for the global metrics and 0.70 ± 0.17 for the regional gradient scores (Supplementary Figs. 7, 8 and Supplementary Tables 18–20). The differences in both gradient range and variation between MDD subgroups remained significant while considering different validation strategies (Supplementary Table 21).

Discussion

In this study, we for the first time report connectome gradient dysfunction in patients with MDD and its association with transcriptional profiles and treatment outcomes. These findings advance our understanding of the neurobiological mechanisms underlying sensory-cognitive deficits and provide potential biomarkers for treatment evaluations in MDD.

In patients with MDD, the lower explained ratio of the principal gradient represents less connectivity variance in the functional connectome, while the narrower range and smaller variance indicate a less differentiated connectivity pattern between the primary and transmodal areas. The DMN regions are geometrically located in the cortex at maximal spatial distances from the primary system, corresponding to their distribution in the embedded gradient axis [17]. The maximal differences in terms of the physical distance and connectivity dissimilarity ensure the complete processing route from mapping concrete stimuli to integrating abstract conceptions [52] while increasing divergence between the primary and transmodal areas to facilitate the formation of abstract cognition by avoiding the interference of input noise [53]. Our meta-analysis confirms that MDD-related alterations in the focal gradient are located in brain areas involving multidomain cognitive functions, including sensory and perceptual processes and higher-order functions. These functions have been reported to be impaired in patients with MDD [2,3,4,5,6,7,8]. Thus, our results provide a potential linkage between connectome gradient disruption and concurrence of sensory-cognitive impairments in MDD. We also show that the topography alteration of the principal gradient was correlated with reduced functional segregation (in terms of clustering coefficient) of the brain network in MDD. This reduction in functional segregation represents a shift toward a randomized network configuration, which has been observed in previous studies of MDD [22, 54, 55]. Here, our results established the connections between alterations of the network gradient and topology in MDD. Importantly, the principal gradient scores of the DMN contributed greatly to our prediction model for antidepressant outcomes. This finding is comparable with several previous reports, demonstrating that the DMN distinguishes remitters from non-remitters in antidepressant medication [56] and predicts patients’ response to sertraline [57]. Our results extended this knowledge that not only the DMN connectivity itself but also its differentiated patterns to other systems are potential markers for clinical outcomes in depression. Notably, the principal gradient explained on average 12% of the connectome variance at the individual level in the current study. This ratio is relatively low compared to gradients decomposed from a group-level matrix [17], however, it is comparable to some individual-level analyses [58].

Our connectome-transcriptome association analysis established a link between MDD-related changes in connectome gradients and gene expression enriched in transsynaptic signaling and calcium ion binding. Transsynaptic signaling is among the most fundamental biological processes that contributes to a series of critical molecular functions, including instructing the formation of synapses and regulating synaptic plasticity [59, 60]. Disruptions in transsynaptic signaling in many key pathways can influence the formation and stability of synapses and play critical roles in the pathology of depression [61]. Studies involving postmortem tissues and rodent models suggest that exposure to chronic stress disrupts the pathway of brain-derived neurotrophic factor (BDNF)-tropomyosin-related kinase B (TrkB) receptor signaling by reducing the downstream extracellular signal-regulated kinase (ERK) and Akt pathways in the hippocampus and prefrontal cortex [62, 63]. Disturbances in these signaling pathways can decrease the expression and function of BDNF and further cause neuronal atrophy in regions implicated in depression [64]. In addition to transsynaptic signaling, calcium ion binding is a molecular function crucial for intracellular signaling. Calcium ion binding can occur in signal transduction resulting from the activation of ion channels or as a second messenger in wide-ranging physiological pathways involving synaptic plasticity. Evidence from postmortem studies suggests that the density of calbindin-immunoreactive GABAergic neurons is reduced in the dorsolateral prefrontal cortex of MDD patients [65]. Currently, measuring regional gene expression in the brain in vivo is extremely difficult. Thus, our results offer a preliminary clue in explaining the relationship between these microscale biological events and macroscopic network alterations in MDD.

Limitations and future research

First, the current work is a retrospective study in which the cognitive performance of the patients was not measured. Alternatively, we examined the associations between the gradient alteration maps and meta-analytic cognitive terms from the Neurosynth database [45]. These results should be considered as evidence of an indirect brain-cognition association in MDD. Second, all patients involved in the prediction analysis were responders to escitalopram, given that patients who had a poor response discontinued the medication or changed their treatment plans. Future studies need to include more non-responders to establish prediction models for treatment‐resistant depression. Third, several critical variables related to the pathology of MDD, such as poor sleep, social stigma, and chronic stress, were not collected in the current dataset. These variables should be recorded in future studies to assess their effects on the functional connectome in MDD. Fourth, the gene expression data from the AIBS were sampled from donors without a diagnosis of MDD, and were susceptible to intersubject variability. A larger sample of whole-brain gene expression data from patients with MDD is required to further characterize and validate the connectome-transcriptome relationship. Fifth, this study applied the ComBat method to harmonize gradient measures across sites. However, potential nonlinear and site-by-group interactive effects remain to be elucidated, for example using traveling-subject datasets and nonlinear models. Finally, the observed effect size of the between-group gradient differences is relatively small (d = 0.16–0.21) but is comparable with previous large-sample imaging studies of MDD from the ENIGMA project [11, 66] and UK Biobank [67]. The small effect size might be due to the high heterogeneity among the MDD patients. Using multivariable statistical methods and considering the subtypes of MDD could provide more effective strategies for improving the statistical power.