Introduction
Major depressive disorder (MDD) is a pervasive mental health condition characterized by persistent feelings of sadness, loss of interest, and a range of cognitive and somatic symptoms (Marx et al., Reference Marx, Penninx, Solmi, Furukawa, Firth, Carvalho and Berk2023). It is one of the leading causes of disability worldwide and poses a significant burden on affected individuals, families, and healthcare systems (Walker, McGee, & Druss, Reference Walker, McGee and Druss2015). While MDD can occur at any age, adolescence represents a particularly vulnerable developmental period (Thapar, Eyre, Patel, & Brent, Reference Thapar, Eyre, Patel and Brent2022). During this critical window, the brain undergoes extensive structural and functional maturation, making it highly susceptible to neurobiological perturbations that may predispose individuals to depression (Paus, Keshavan, & Giedd, Reference Paus, Keshavan and Giedd2008; Whittle et al., Reference Whittle, Lichter, Dennison, Vijayakumar, Schwartz, Byrne, Simmons, Yücel, Pantelis, McGorry and Allen2014). Understanding the neurobiological underpinnings of adolescent MDD is therefore essential for developing effective early interventions and preventive strategies.
Using multimodal magnetic resonance imaging (MRI), previous studies have well demonstrated that adolescent MDD is associated with regional structural and functional brain abnormalities, predominantly in regions involved in emotional regulation and cognitive processing, such as the prefrontal cortex, anterior cingulate, and limbic structures (Long et al., Reference Long, Li, Wang, Cao, Wu, Roberts, Gong, Kemp and Jia2023; Wu et al., Reference Wu, Zhang, Xie, Zhang, Ling, Gan, Qiu, Roberts, Jia and Gong2025; Zhang et al., Reference Zhang, Wu, Zhang, Xie, Ling, Zhao, Gan, Qiu, Mechelli, Jia and Gong2025). Based on brain connectomes derived from structural or functional MRI, recent studies have highlighted network-level abnormalities in adolescents with MDD (Tse et al., Reference Tse, Ratheesh, Ganesan, Zalesky and Cash2023; Wu et al., Reference Wu, Li, Zhou, Zhang and Long2020; Reference Wu, Chen, Long, Cao, Xie, Wang, Roberts, Gong and Jia2023, Reference Wu, Long, Cao, Xie, Wang, Roberts, Gong and Jia2024), providing a more comprehensive understanding of the neural mechanisms underlying the disorder. Particularly, the structural similarity network, derived from high-resolution T1-weighted anatomical images, has emerged as a powerful tool for capturing the intricate organization of cortical features and identifying subtle structural alterations associated with MDD. As a more novel network model, the Morphometric Inverse Divergence (MIND) has been proposed to better characterize inter-regional structural similarity based on the divergence between multivariate distributions of multiple structural MRI features (Sebenius et al., Reference Sebenius, Seidlitz, Warrier, Bethlehem, Alexander-Bloch, Mallard, Garcia, Bullmore and Morgan2023). For neuroimaging studies, the MIND network model has several key advantages compared to traditional morphological similarity network (MSN). First, it yields more reliable and robust mapping of individual structural connectomes, since the structural similarity-based connectomes produced by MIND align more closely with structural similarity principles (Sebenius et al., Reference Sebenius, Seidlitz, Warrier, Bethlehem, Alexander-Bloch, Mallard, Garcia, Bullmore and Morgan2023). Second, the MIND network appears to be more heritable, showing stronger associations with cortical gene co-expression, particularly for connections linking structurally differentiated regions (Sebenius et al., Reference Sebenius, Seidlitz, Warrier, Bethlehem, Alexander-Bloch, Mallard, Garcia, Bullmore and Morgan2023). Collectively, these advantages of the MIND network model suggest that it could be more effective for studying the cortical structural connectomes in MDD. Using the MIND method, a recent study has established a normative framework of cortical morphometric networks and found substantial individual heterogeneity in network deviations in MDD (Liang et al., Reference Liang, Sun, Xia, Zhao, Gong, Li, Liao, Cui, Duan, Pang, Wang, Yu, Bi, Chen, Chen, Chen, Chen, Cheng, Cheng and He2025). However, studies investigating cortical morphometric network patterns in adolescent patient samples are still limited.
Although advancements in neuroimaging have revealed the neural substrates of adolescent MDD, the underlying biological mechanisms remain unclear. Therefore, combining macroscale imaging representations with microscale gene expression data has become an important research frontier in this field. Emerging evidence suggests that genes play a crucial role in organizing human brain networks (Arnatkeviciute et al., Reference Arnatkeviciute, Fulcher, Oldham, Tiego, Paquola, Gerring, Aquino, Hawi, Johnson, Ball, Klein, Deco, Franke, Bellgrove and Fornito2021). Moreover, across development from prenatal to adolescence, genetic abnormalities can alter normal brain development and lead to psychotic symptoms in various disorders (Rapoport, Giedd, & Gogtay, Reference Rapoport, Giedd and Gogtay2012). Connectome–transcriptome association analysis based on the Allen Human Brain Atlas (AHBA) database can now investigate the spatial correlations between microscale gene expression profiles and macroscale network changes, providing deep insights into the molecular basis underlying neural abnormalities (Arnatkeviciute et al., Reference Arnatkeviciute, Markello, Fulcher, Misic and Fornito2023). Using AHBA gene expression data, recent connectome−transcriptome association studies have identified transcriptomic signatures that underlie neuroimaging changes in various neuropsychiatric and neurodevelopmental disorders (Buch et al., Reference Buch, Vértes, Seidlitz, Kim, Grosenick and Liston2023; Li et al., Reference Li, Seidlitz, Suckling, Fan, Ji, Meng, Yang, Wang, Qiu, Chen and Liao2021; Qin et al., Reference Qin, Li, Zhang, Yin, Wu, Pan, Chen, Roberts, Sweeney, Huang, Gong and Jia2024). However, the potential molecular mechanisms underlying cortical structural network abnormalities in adolescents with MDD remain unclear.
Thus, to address these knowledge gaps, this study aims to elucidate the cortical structural network abnormality patterns in adolescents with MDD using a novel MIND network model. Additional machine learning models were developed to evaluate the potential of MIND measures as clinical supplementary biomarkers for diagnosing adolescent MDD. Furthermore, connectome–transcriptome correlation analysis was performed to explore the spatial associations between MIND changes and transcriptional profiles using the brainwide gene expression data (i.e. AHBA atlas). To gain a deep understanding of the molecular mechanisms underlying MIND abnormalities in adolescent MDD, we employed various gene expression enrichment analyses. We hypothesized that (1) adolescent MDD would show MIND changes in specific regions and networks compared to healthy controls (HCs); (2) adolescent MDD-related regional MIND alterations would be linked to specific biological pathways and cell types; and (3) machine learning models leveraging regional MIND features would achieve individual-level classification of adolescents with MDD versus HCs.
Methods
Participants
This study received approval from the Ethics Committee of West China Hospital, Sichuan University. Written informed consent was obtained from all participants and their guardians before the study. This study recruited 178 adolescents with MDD and 82 HCs who were right-handed and aged 12–18 years. All patients were clinically assessed and diagnosed according to the DSM-5 criteria by trained psychiatric clinicians. The severity of depression was assessed using the 17-item Hamilton Rating Scale for Depression (HAMD-17). Patients were eligible if they were experiencing their first episode of illness, had no prior treatment with antidepressants, antipsychotics, or other psychotropic drugs, scored at least 17 on the HAMD-17, and had no comorbid psychiatric diagnoses such as bipolar disorder, schizophrenia, or anxiety disorders. Healthy adolescents were recruited from local communities or schools and were included if they and their first-degree relatives had no history of psychiatric disorders. Exclusion criteria for all participants were: (i) the presence of neurological disorders, significant medical illnesses, or developmental disorders; (ii) a history of substance abuse or dependence; (iii) contraindications to MRI scanning (e.g. metal implants and claustrophobia); and (iv) poor image quality. An overview of the analytical framework is summarized in Supplementary Figure S1.
Structural MRI acquisition and data preprocessing
High-resolution structural MRI data were acquired on a 3.0T scanner (GE Discovery MR750) using a 3D-BRAVO sequence: sagittal 188 slices covering the whole brain; repetition time, 8.2 ms; echo time, 3.2 ms; field of view, 256 × 256 mm2; matrix size, 256 × 256; voxel size, 1.0 × 1.0 × 1.0 mm3; flip angle, 12°; slice thickness, 1.0 mm, without a gap. Subsequently, standard data preprocessing was performed using FreeSurfer 7.0 (http://surfer.nmr.mgh.harvard.edu/). The preprocessing steps included skull stripping, tissue segmentation, surface reconstruction, metric reconstruction, and spherical normalization parameter estimation. Additionally, we calculated the total intracranial volume (TIV) for each subject. Quality control procedures were implemented to ensure higher image quality using the CAT12 toolbox (http://dbm.neuro.uni-jena.de/cat12/). Image quality was quantified using the Image Quality Rating (IQR). We applied a stricter cutoff by excluding participants with IQR < 80 from the analyses. As a result, four patients were excluded.
MIND network construction
According to the Desikan-Killiany (DK-68) atlas, the cortical surfaces were subdivided into 308 spatially contiguous regions of interest (ROIs) via a backtracking algorithm. Using inverse spherical normalization parameters estimated in cortical surface reconstruction, this parcellated DK-308 atlas was then mapped from standard space into the native cortical space of each participant. Subsequently, the gray matter volume, cortical thickness, surface area, mean curvature, and sulcal depth were extracted for all vertices within each ROI based on the above transformation (Sebenius et al., Reference Sebenius, Seidlitz, Warrier, Bethlehem, Alexander-Bloch, Mallard, Garcia, Bullmore and Morgan2023). The five extracted morphometric features underwent z-score normalization across all vertices in the brain to standardize their scales, generating regional multivariate distributions. Using a transformed symmetrized Kullback–Leibler divergence, we calculated pairwise similarities among these distributions, with values bounded between 0 and 1. As a result, a 308 × 308 cortical morphometric similarity matrix was generated for each subject, with higher similarity values indicating greater similarity, and vice versa. Finally, regional MIND strength, defined as the weighted nodal degree, was derived from the similarity matrix by averaging the edge weights for each of the 308 cortical regions, with no thresholding applied. Furthermore, we calculated a global MIND metric as the mean of regional MIND values across the 308 regions to compare global MIND strength between adolescents with MDD and HCs.
Group differences in regional MIND
A general linear model was used to analyze alterations of regional MIND in adolescents with MDD compared to HCs, with age, sex, and TIV as covariates. Subsequently, case–control t-values of regional MIND were obtained. To further examine MIND changes at the network level in adolescent MDD patients, we employed the cellular structure von Economo atlas (Cauda et al., Reference Cauda, Torta, Sacco, D’Agata, Geda, Duca, Geminiani and Vercelli2013) to partition the 308 regions into seven subnetworks. Then we calculated the differences in MIND values between the two groups at various subnetwork levels, using the same statistical framework as mentioned above. The statistical level was set at p < 0.05 after false discovery rate (FDR) correction. Spearman correlation analyses were performed to assess the relationships between significantly altered regional MIND and clinical features within the adolescent MDD group.
Machine learning classification model
To assess the potential of regional MIND as diagnostic biomarkers for adolescent MDD, a machine learning classification model was developed to distinguish between adolescents with MDD and HCs. Models were constructed with and without feature selection to determine the optimal model. Feature selection was performed using a two-sample t-test, retaining features that demonstrated statistical significance at p < 0.05 (uncorrected). Subsequently, multilayer perceptron, k-nearest neighbor, logistic regression, support vector machine (SVM), extreme gradient boosting, and random forest were used to construct classification models. A 10-fold cross-validation was applied to evaluate model performance. Considering class imbalance, we employed SMOTE on the training set to synthetically augment minority class cases and improve model learning performance. External validation analysis was performed to assess the generalizability of the optimal model (Supplementary Methods). To assess whether MIND outperformed traditional MSN in the diagnosis of adolescent MDD, we calculated individual MSN and constructed MSN-based models for comparison (Supplementary Methods). Model performance was evaluated using accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and the area under the receiver operating characteristic curve (AUC).
Furthermore, the SHapley Additive exPlanations (SHAP) algorithm was utilized to interpret the contribution of different features to the model’s predictions. One important advantage of the SHAP method is its capability to show the significance of each feature for individual predictions, as well as the overall average importance across the entire dataset (Lundberg et al., Reference Lundberg, Erion, Chen, DeGrave, Prutkin, Nair, Katz, Himmelfarb, Bansal and Lee2020). Thus, this technique provides insights into how and why a model reaches a particular decision, enabling clinicians to gain a better understanding of the contributing factors behind predictions and thereby enhancing the clinical interpretability of the model outcomes. We computed SHAP values for the factors in the optimal model. Positive SHAP values suggest that a feature contributes positively to the prediction, whereas negative values indicate a detrimental effect. Additionally, all factors were normalized with their absolute values to compare the relative importance (Wang et al., Reference Wang, Lin, Zhu, Chen, Zou, Wang, Huo and Zhou2025).
Connectome–transcriptome association analysis
Gene expression data from six healthy adult donors were acquired from the AHBA atlas (http://human.brain-map.org/). Since only two donor brains in the AHBA dataset provided samples of the right hemisphere, this study only analyzed the gene expression data of the left hemisphere. The abagen (https://github.com/rmarkello/abagen) toolbox was used for preprocessing the AHBA data according to the recommended pipeline of connectome-transcriptome analysis (Arnatkeviciute, Fulcher, & Fornito, Reference Arnatkeviciute, Fulcher and Fornito2019). The preprocessing steps included reannotating probes to genes, filtering probes, selecting representative probes, assigning samples to brain regions, and normalizing the expression data using the z-score method. By averaging the samples from all six donors, gene expression levels were calculated for each of the 152 regions, generating a gene expression matrix (152 regions × 15632 genes).
Next, a connectome–transcriptome association analysis was performed using partial least squares (PLS) regression, with the case–control t-statistic map of the regional MIND differences as the response variable and the normalized gene expression matrix as the predictor variable. The fitted PLS regression model can produce multiple components, each representing a linear combination of the predictor variables that explains variance in the response variables. The first or second component (PLS1 or PLS2) was selected, as it usually accounts for the largest proportion of variance in the response. The significance of the variance explained by PLS1 or PLS2 was evaluated using a spatial permutation test (10,000 iterations) that incorporated correction for spatial autocorrelation. Subsequently, a bootstrapping (1,000 iterations) method was employed to estimate and adjust the error in estimating the weight of each gene’s contribution to the PLS model. The corrected gene weights were then determined by the ratio of the raw weight to the estimated bootstrapping error. The genes were ranked based on their corrected weights in the PLS model. Finally, separate gene ranking lists were obtained for genes with significant positive (PLS+) and negative (PLS−) weights (p < 0.01, FDR-corrected) (Xue et al., Reference Xue, Guo, Zhu, Liang, Xu, Ma, Liu, Zhang and Liu2023).
Enrichment analysis
To evaluate whether the significant PLS-weighted gene enriched in specific biological pathways, cell types, and developmental windows, we performed a series of enrichment analyses on the identified PLS+ and PLS− gene lists. First, functional enrichment analyses were performed using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG), implemented via the Metascape platform (Zhou et al., Reference Zhou, Zhou, Pache, Chang, Khodabakhshi, Tanaseichuk, Benner and Chanda2019). The enrichments were tested at a statistical significance of p < 0.05 after FDR correction. Then, cell-type-specific expression analysis (CSEA) (http://genetics.wustl.edu/jdlab/csea-tool-2) was conducted to assess the cell specificity of the PLS-weighted genes. Initially, we employed the CSEA to preliminarily investigate specific cell types related to both PLS+ and PLS− genes. To explore the potential cellular origins of the genes, we overlapped the PLS+ and PLS− genes with gene sets of seven major brain cortical cell types (excitatory neurons, inhibitory neurons, oligodendrocytes, astrocytes, microglia, endothelial cells, and ependymal cells) (Li et al., Reference Li, Seidlitz, Suckling, Fan, Ji, Meng, Yang, Wang, Qiu, Chen and Liao2021). For each cell type, we counted the number of overlapping genes in both PLS+ and PLS− genes. The statistical significance of these overlaps was assessed using a permutation test with 10,000 iterations and adjusted for multiple testing with FDR (p < 0.05, FDR-corrected). To highlight developmental time windows across brain regions where genes are strongly expressed, we also conducted developmental gene expression enrichment analysis using the CSEA developmental express tool, which is based on spatiotemporal-transcriptomic data from the BrainSpan dataset (http://www.brainspan.org). Statistical significance was set at p < 0.05 after FDR correction.
Results
Demographic and clinical characteristics
This study finally included 174 adolescents with MDD and 82 HCs. The two groups were matched in age, gender, and education. The detailed demographic and clinical characteristics are shown in Table 1.
Demographic and clinical characteristics of the participants
Table 1. Long description
The table consists of four columns: Characteristic, Adolescent M D D patients (n = 174), Healthy controls (n = 82), and p-value.
* Gender (male/female): M D D group is 83/91; Healthy controls are 36/46; p-value is 0.570.
* Age (years): M D D group is 15.74 plus or minus 1.70; Healthy controls are 16.10 plus or minus 1.64; p-value is 0.109.
* Education (years): M D D group is 8.82 plus or minus 1.72; Healthy controls are 9.11 plus or minus 1.66; p-value is 0.200.
* B M I (kg/m super 2): M D D group is 21.42 plus or minus 2.18; Healthy controls are 21.44 plus or minus 1.74; p-value is 0.941.
* Disease duration (months): M D D group is 10.78 plus or minus 7.19; Healthy controls and p-value are not applicable.
* Age of onset (years): M D D group is 14.97 plus or minus 1.82; Healthy controls and p-value are not applicable.
* H A M D score: M D D group is 23.47 plus or minus 4.56; Healthy controls are 2.52 plus or minus 1.15; p-value is less than 0.001.
Quantitative data are expressed as mean plus or minus standard deviation. M D D stands for major depressive disorder; B M I stands for body mass index; H A M D stands for Hamilton Depression Rating Scale.
Note: All quantitative data are expressed as mean ± standard deviation; numbers for gender data. MDD, major depressive disorder; BMI, body mass index; HAMD, Hamilton Depression Rating Scale.
a The p-value was calculated by using χ 2-test.
b The p-value was calculated by using independent two-sample t-test.
Regional MIND alterations in adolescents with MDD
No significant difference was found in global MIND value between the two groups (p = 0.718). The two groups showed similar distribution of MIND strength across the whole brain (Figure 1a). However, we found significant differences in regional MIND between the two groups. Specifically, adolescents with MDD showed significantly lower regional MIND in the bilateral pericalcarine cortex (part1), left precentral gyrus (part3), bilateral lingual gyrus (part2), right precentral gyrus (part6), right pericalcarine cortex (part2), and right lingual gyrus (part6), as well as significantly higher regional MIND in the right lateral orbitofrontal cortex (part4), compared to HCs (Figure 1b and 1c). The mean MIND strength of the HC group showed a positive spatial correlation with the case–control t-value (r = 0.152, p spin = 0.007) (Figure 1d).
MIND change patterns and their relationships to clinical variables in adolescent major depressive disorder (aMDD) patients. (a) aMDD patients and healthy controls (HCs) showed similar patterns of regional MIND distributions. (b) Regions showing significant differences in regional MIND between aMDD patients and HCs. (c) Box plots displaying differences in regional MIND values for the statistically significant regions between the two groups. (d) The mean MIND strength of the HC group exhibited a positive spatial correlation with the case–control t-values (r = 0.152, p spin = 0.007). (e) Regional MIND in the left precentral gyrus negatively correlated with age of onset in aMDD patients (r = −0.230, p = 0.002). (f) Regional MIND in the right lateral orbitofrontal cortex negatively correlated with (r = −0.153, p = 0.044) age of onset in aMDD patients. *p < 0.05, **p < 0.01 (FDR-corrected).
Figure 1. Long description
Panel A presents brain surface maps for Mean a M D D M I N D and Mean H C M I N D, showing similar regional distributions on a color scale from 0.0 blue to 0.16 red.
Panel B displays Case-control t-maps and Statistically significant regions on brain surfaces, with t-values ranging from -4.5 blue to 4.5 red.
Panel C is a box plot comparing Regional M I N D values between a M D D (red) and H C (green) across nine brain regions: lh_pericalcarine_part1, lh_precentral_part3, rh_lingual_part2, rh_precentral_part6, rh_pericalcarine_part2, rh_lateralorbitofrontal_part4, rh_pericalcarine_part1, lh_lingual_part2, and rh_lingual_part6. Significant differences are marked with asterisks.
Panel D is a scatter plot showing a positive spatial correlation between Mean H C M I N D (x-axis) and Case-control t-value (y-axis), with r = 0.152 and p spin = 0.007.
Panel E is a scatter plot showing a negative correlation between Age of onset in years (x-axis) and M I N D in left precentral gyrus (y-axis), with r = -0.230 and p = 0.002.
Panel F is a scatter plot showing a negative correlation between Age of onset in years (x-axis) and M I N D in right lateral O F C (y-axis), with r = -0.153 and p = 0.044.
We also analyzed the network-level MIND differences between adolescents with MDD and HCs using a predefined atlas. At the von Economo atlas level, adolescents with MDD had significantly lower MIND in the primary motor and primary sensory networks, and significantly higher MIND in the limbic network (p < 0.05, FDR-corrected) compared to HCs (Supplementary Figure S2).
Furthermore, regional MIND values in the left precentral gyrus (r = −0.230, p = 0.002, uncorrected) (Figure 1e) and the right lateral orbitofrontal cortex (r = −0.153, p = 0.044, uncorrected) (Figure 1f) were negatively associated with age of onset in adolescents with MDD. No significant correlations were found between MIND changes and other clinical variables, including illness duration and HAMD scores.
Performance of the regional MIND-based classification model
After feature screening via t-test (p < 0.05, uncorrected), MIND values of 38 brain regions were selected as the input features. Notably, for each model, feature selection led to better performance compared to models without feature selection (Supplementary Table S1). The SVM model with feature selection showed the best performance (AUC = 0.824) and generalized well to the external independent dataset (AUC = 0.724) (Figure 2a and 2b; Supplementary Table S2). Moreover, MIND-based models outperformed MSN-based models in distinguishing adolescent MDD patients from HCs (Supplementary Table S3).
Performance and feature interpretation of the optimal machine learning model. (a) Confusion matrix for the testing set of the optimal SVM model. (b) The receiver operating characteristic curve (ROC) showing the AUCs of the optimal SVM model in the training, testing, and external validation sets. (c) SHAP summary plot of features included in the classification model. (d) The mean absolute SHAP value summary plot, representing the average impact of each feature on the classification of aMDD. Note: aMDD, ‘adolescent major depressive disorder’; HC, ‘healthy control’; AUC, ‘area under the curve’; SVM, ‘support vector machine’.
Figure 2. Long description
The figure consists of four panels labeled A through D.
* Panel A: Confusion matrix for the testing set. The y-axis is True class and the x-axis is Predicted class. Top-left quadrant (H C, H C) shows 0.692 True negative. Top-right (H C, a M D D) shows 0.308 False positive. Bottom-left (a M D D, H C) shows 0.103 False negative. Bottom-right (a M D D, a M D D) shows 0.897 True positive. A color scale on the right ranges from 0 (white) to 1 (dark blue).
* Panel B: R O C curves plotting Sensitivity on the y-axis against 1 minus Specificity on the x-axis. Three stepped lines represent the Training set (red, A U C 0.904), Testing set (blue, A U C 0.824), and Validation set (black, A U C 0.724). All curves rise sharply from the origin toward the top-left corner.
* Panel C: S H A P summary plot. The y-axis lists anatomical brain features (e.g., rh_lateralorbitofrontal_part4). The x-axis shows S H A P value from negative 0.075 to 0.100. Each row contains a distribution of colored dots where yellow represents high feature value and purple represents low feature value. Dots spread horizontally around a vertical zero line.
* Panel D: Mean absolute S H A P value summary plot. A horizontal bar chart ranking features by their average impact. The top feature is rh_lateralorbitofrontal_part4 with a value of 0.0265. The bars decrease in length as they move down the y-axis, ending with lh_postcentral_part1 at 0.0042. The bars are color-coded in a gradient from yellow at the top to dark purple at the bottom.
In addition, we used the SHAP analysis to interpret the importance of individual features in the optimal classification model. The top 10 predictive MIND features were mainly located at the lateral orbitofrontal cortex, middle frontal cortex, superior frontal cortex, and anterior cingulate cortex (Figure 2c and 2d).
Transcriptional patterns associated with regional MIND changes
The PLS1 explained the highest variation (17.02%) in MIND differences between adolescents with MDD and HCs (p < 0.05) (Supplementary Figure S3). Interestingly, the PLS1-weighted gene expression map showed a distribution pattern characterized by a gradual increase in gene expression from anterior to posterior brain regions (Figure 3a). The PLS1 score showed a significant positive spatial correlation with the case–control t-map of regional MIND differences (Figure 3b) (r = 0.334, p spin < 0.001), suggesting that the PLS1-weighted genes were simultaneously overexpressed in regions showing increased regional MIND in adolescent MDD patients. Moreover, we identified significant positive (651 genes) and negative (846 genes) PLS1 ranking gene lists using the univariate one-sample Z test (p < 0.01, FDR-corrected) (Figure 3c and 3d).
Transcriptional profiles spatially associated with case–control t-values of regional MIND changes. (a) The regional MIND change pattern (case–control t-map) coincided with the left-hemisphere PLS1-weighted gene expression map. (b) Case–control t-values of regional MIND changes showed a positive spatial correlation with the PLS1 scores (r = 0.334, p spin < 0.001). (c) According to the ranked Z scores, 651 PLS1+ genes and 846 PLS1− genes were identified as significant (p < 0.01, FDR-corrected). (d) Lists of the PLS1+ and PLS1− genes.
Figure 3. Long description
Panel A shows four brain surface maps of the left hemisphere. The top pair displays the Case-control t-map with a color scale from blue (t-value -4.5) to red (t-value 4.5). The bottom pair displays the Weighted gene expression map (P L S 1) with a color scale from blue (P L S 1 score -0.25) to red (0.15). The spatial patterns of red and blue regions appear highly similar between the two sets.
Panel B is a scatter plot with P L S 1 scores on the x-axis (ranging from -0.25 to 0.15) and Case-control t-values on the y-axis (ranging from -5 to 3). Red data points show a positive linear trend, indicated by a red regression line. Text in the bottom right corner states r equals 0.334, p sub spin is less than 0.001.
Panel C is a histogram titled Relative counts on the y-axis (0 to 1200) and P L S 1 weighted Z score on the x-axis (-6.00 to 6.00). Vertical orange lines mark the p F D R less than 0.01 threshold at Z scores of -3.31 and 3.31. The left tail (teal) identifies P L S 1 minus genes (n equals 846) and the right tail (pink) identifies P L S 1 plus genes (n equals 651).
Panel D is a table of Gene symbols and Loadings. The top section (P L S 1 plus) lists genes in a yellow-to-green gradient: C X orf 5 7 (7.180), C 2 C D 4 C (6.877), C 1 Q L 3 (6.800), B A I A P 3 (6.633), and N E C A B 2 (6.608). The bottom section (P L S 1 minus) lists genes in a blue-to-purple gradient: R P G R (-6.635), S T 3 G A L 6 (-6.762), T I F A (-6.865), T R P C 3 (-7.443), and S Y N E 4 (-7.630).
Functional pathways associated with regional changes in MIND
GO and KEGG enrichment analysis revealed that PLS1+ genes showed prominent enrichment in synapse signaling-related pathways and neural development-related biological processes (Figure 4a and 4b; Supplementary Figure S4A). PLS1− genes were predominantly enriched in pathways and biological processes related to metabolism and transport (Figure 4c and 4d; Supplementary Figure S4B).
Functional enrichment analysis of the significant PLS1-weighted genes. (a) Bubble diagram displaying the enrichment of GO terms for the PLS1+ genes. Bubble size represents the number of genes assigned to each GO term. The color bar represents the significance level. (b) Enrichment network showing intra- and inter-cluster similarities of the enriched GO terms for the PLS1+ genes. (c) Bubble diagram displaying the enrichment of GO terms for the PLS1− genes. Bubble size represents the number of genes assigned to each GO term. The color bar represents the significance level. (d) Enrichment network showing intra- and inter-cluster similarities of the enriched GO terms for PLS1− genes.
Figure 4. Long description
Panel A is a bubble diagram for P L S 1 plus genes. The y-axis lists G O terms like dendrite and presynapse. The x-axis is GeneRatio percentage from 2 to 8. Bubble size indicates gene count from 20 to 50. Color indicates minus log 10 p F D R from 6 in red to 14 in blue. Dendrite shows the highest GeneRatio and significance.
Panel B is an enrichment network for P L S 1 plus genes. A dense central cluster includes nodes for dendrite, modulation of chemical synaptic transmission, and monoatomic ion channel activity. Peripheral clusters include regulation of system process at the top and kinase regulator activity at the bottom.
Panel C is a bubble diagram for P L S 1 minus genes. The y-axis lists terms like metal ion transport and active transmembrane transporter activity. The x-axis is GeneRatio from 0 to 6. Significance ranges from 3 in red to 8 in blue. Metal ion transport has the highest GeneRatio and significance.
Panel D is an enrichment network for P L S 1 minus genes. It features a central chain of clusters including protein complex oligomerization and metal ion transport. Distinct peripheral clusters include neuronal cell body to the left, mitochondrial membrane at the top right, and protein homodimerization activity at the bottom right.
Cellular signatures associated with regional changes in MIND
The CSEA revealed distinct patterns for PLS1+ and PLS1− genes. In the PLS1+ gene set, significant enrichment was observed in neurons and astrocytes (p < 0.05, FDR-corrected) (Figure 5a). PLS1− genes also exhibited significant enrichment mainly in neurons (p < 0.05, FDR-corrected) (Figure 5b). Moreover, we found significant overlaps between the PLS1+ genes and excitatory neurons, inhibitory neurons, and astrocytes (Figure 5c), while the PLS1− genes showed significant overlaps with the excitatory neurons, inhibitory neurons, and endothelial cells (Figure 5d) (p < 0.05, FDR-corrected).
Cellular-level transcriptional signatures associated with MIND changes in adolescents with major depressive disorder. (a) Cell-type-specific expression analysis (CSEA) revealed significant enrichment of the PLS1+ genes in neurons and astrocytes. (b) CSEA revealed significant enrichment of the PLS1− genes in neurons. (c) The number of PLS1+ genes overlapped with gene sets for each cell type. (d) The number of PLS1− genes overlapped with gene sets for each cell type. An asterisk indicates significant overlaps by permutation tests (FDR-corrected).
Figure 5. Long description
Panel A and B are radial dendrograms where nodes represent specific cell types and brain regions. Each node is a concentric hexagon. A color scale at the bottom of each panel indicates p-values from 0.1 in yellow to 0 in dark red.
Panel A shows P L S 1 plus genes. Significant dark red enrichment is visible in several nodes including Layer 5 b Neurons in the Cortex, Layer 6 Neurons in the Cortex, Astrocytes in the Cortex, and Bergmann Glia and Oligodendrocytes in the Cerebellum.
Panel B shows P L S 1 minus genes. Most nodes are gray, indicating no significance, but significant enrichment is shown in Cholinergic Motor Neurons in the Spinal Cord and Brain Stem, and Purkinje Neurons in the Cerebellum.
Panel C and D are horizontal bar charts with Cell types on the y-axis and Number of overlapped genes on the x-axis.
Panel C for P L S 1 plus genes shows the following order from top to bottom:
* Astro: approximately 95 genes with an asterisk.
* Neuro-Ex: approximately 94 genes with an asterisk.
* Neuro-In: approximately 75 genes with an asterisk.
* Endo: approximately 32 genes.
* Micro: approximately 24 genes.
* Oligo: approximately 18 genes.
* O P C: approximately 10 genes.
Panel D for P L S 1 minus genes shows the following order from top to bottom:
* Endo: approximately 74 genes with an asterisk.
* Neuro-Ex: approximately 71 genes with an asterisk.
* Neuro-In: approximately 56 genes with an asterisk.
* Micro: approximately 34 genes.
* Astro: approximately 27 genes.
* Oligo: approximately 15 genes.
* O P C: approximately 3 genes.
Developmental enrichments associated with regional changes in MIND
The developmental gene expression enrichment analysis revealed critical spatiotemporal windows for gene expression. Notably, the PLS1+ genes showed predominant expression across cortex, striatum, hippocampus, and amygdala from early/mid fetal to adolescence. In contrast, the PLS1− genes exhibited significant enrichment in the cerebellum and thalamus during infancy and young adulthood (Supplementary Figure S5).
Discussion
By integrating neuroimaging and gene expression data, the present study provides an in-depth investigation of macroscale cortical structural abnormalities and the associated microscale molecular mechanisms in adolescents with MDD. This study showed that regional MIND abnormalities in adolescent MDD patients were primarily located in the orbitofrontal cortex, as well as in sensorimotor and visual network regions. Additionally, the developed machine learning model demonstrated the value of MIND networks in the classification of adolescents with MDD versus HCs. Furthermore, we identified that adolescent MDD-related changes in regional MIND were associated with specific biological pathways and cell types. These findings yield novel insights into the neurobiology of adolescent MDD.
Our findings demonstrated that adolescent MDD was associated with regional MIND abnormalities, including decreased MIND in sensory regions and increased MIND in the right lateral orbitofrontal cortex (a paralimbic region). These regional patterns are consistent with prior neuroimaging studies reporting abnormalities in visual and sensorimotor areas, as well as in orbitofrontal regions associated with emotional regulation and decision-making deficits in adolescent MDD (Schmaal et al., Reference Schmaal, Hibar, Sämann, Hall, Baune, Jahanshad, Cheung, van Erp, Bos, Ikram, Vernooij, Niessen, Tiemeier, Hofman, Wittfeld, Grabe, Janowitz, Bülow, Selonke and Veltman2017; Zhang et al., Reference Zhang, Wu, Zhang, Xie, Ling, Zhao, Gan, Qiu, Mechelli, Jia and Gong2025). A recent large-scale lifespan study established that typical MIND development is characterized by divergent trajectories in sensory versus paralimbic cortices (Liang et al., Reference Liang, Sun, Xia, Zhao, Gong, Li, Liao, Cui, Duan, Pang, Wang, Yu, Bi, Chen, Chen, Chen, Chen, Cheng, Cheng and He2025). Specifically, primary sensory cortices exhibit increasing morphometric differentiation (decreasing MIND) from infancy to early adolescence, reflecting a normal process of cortical specialization and synaptic pruning. Conversely, paralimbic regions (e.g. orbitofrontal cortex) show increasing morphometric similarity (increasing MIND) from birth to early adulthood, suggesting a prolonged period of structural plasticity and integration (Liang et al., Reference Liang, Sun, Xia, Zhao, Gong, Li, Liao, Cui, Duan, Pang, Wang, Yu, Bi, Chen, Chen, Chen, Chen, Cheng, Cheng and He2025). Our findings in adolescent MDD revealed a pattern that may represent an exaggeration or derailment of these typical developmental trajectories. The decreased MIND in sensory regions could reflect accelerated or excessive morphometric differentiation beyond the normal developmental range, potentially driven by heightened synaptic pruning or reduced neuroplasticity in these areas during adolescence. In contrast, the increased MIND in the lateral orbitofrontal cortex may indicate abnormally heightened morphometric similarity in a paralimbic region that normally shows strengthening during development. Consistent with previous studies (Pannekoek et al., Reference Pannekoek, van der Werff, Meens, van den Bulk, Jolles, Veer, van Lang, Rombouts, van der Wee and Vermeiren2014; Redlich et al., Reference Redlich, Opel, Bürger, Dohm, Grotegerd, Förster, Zaremba, Meinert, Repple, Enneking, Leehr, Böhnlein, Winters, Froböse, Thrun, Emtmann, Heindel, Kugel, Arolt and Dannlowski2018), this finding particularly underscores the important role of the paralimbic system, which is involved in emotional regulation, in adolescent MDD. Overall, our findings suggest that MDD may involve an exaggeration or derailment of the normal maturational processes during adolescence.
Additionally, our study found a positive spatial correlation between the mean MIND strength in the HC group and the case–control t-values, suggesting that brain regions with higher MIND in normal individuals are more susceptible to MDD-related alterations. This observation is consistent with prior studies showing that brain regions that undergo the most pronounced developmental changes also tend to be the most vulnerable in psychiatric disorders. We also observed weak negative correlations of regional MIND in the left precentral gyrus and the right lateral orbitofrontal cortex with age of onset in adolescents with MDD, suggesting potential interactions between regional MIND in emotional and motor-related regions and developmental trajectories. Moreover, the lack of association with current HAMD severity, together with the observed relationship with age of onset, suggests that the observed MIND alterations are trait-like rather than primarily state-dependent.
This study further elucidated the biological basis of morphometric brain network alterations in adolescent MDD. Gene enrichment analysis suggested that MIND changes in adolescent MDD were associated with synaptic signaling, neurodevelopment, and metabolism-related biological processes, with distinct transcriptomic signatures for different brain regions across key developmental windows. Altered gene expression involved in metabolism, transcription, and neurotransmission has been observed in postmortem brain tissues and is associated with the neurobiology of major depression and suicide (Klempan et al., Reference Klempan, Sequeira, Canetti, Lalovic, Ernst, ffrench-Mullen and Turecki2009; Sequeira et al., Reference Sequeira, Klempan, Canetti, ffrench-Mullen, Benkelfat, Rouleau and Turecki2007). Cellular metabolism and synaptic signaling are important biological processes related to synaptic plasticity (Marsden, Reference Marsden2013). Impaired neuroplasticity may underlie the development of maladaptive responses to stressful experiences (McEwen & Gianaros, Reference McEwen and Gianaros2011). A prior study in adult MDD have revealed that brain morphometric network changes have been associated with similar synaptic signaling pathways (Li et al., Reference Li, Seidlitz, Suckling, Fan, Ji, Meng, Yang, Wang, Qiu, Chen and Liao2021). Moreover, it has been demonstrated that biological processes involved in synaptic signaling, neurodevelopment, and metabolism are important in regulating normative MIND network development (Liang et al., Reference Liang, Sun, Xia, Zhao, Gong, Li, Liao, Cui, Duan, Pang, Wang, Yu, Bi, Chen, Chen, Chen, Chen, Cheng, Cheng and He2025). Dysregulation of these important biological processes may lead to untypical MIND developmental trajectories observed in adolescent MDD. Additionally, the cellular enrichment analysis further clarified these mechanisms by identifying the cell substrates driving morphometric abnormalities. We found significant enrichment of PLS1+ genes in neurons and astrocytes. Astrocytes play a key role in brain circuit development and homeostasis, controlling synapse formation, function, plasticity, and elimination across both adolescent development and adulthood (Fossati, Matteoli, & Menna, Reference Fossati, Matteoli and Menna2020), making them promising targets for antidepressant strategies (Fullana et al., Reference Fullana, Gasull-Camós, Tarrés-Gatius, Castañé, Bortolozzi and Artigas2020; Zhou et al., Reference Zhou, Xiao, Xie, Yang, Wang and Tu2019). Our developmental enrichment analyses revealed that PLS1+ genes showed predominant expression across cortical regions, striatum, and limbic regions during key developmental stages from early/mid fetal to adolescence. These regions parallel the sensorimotor and paralimbic areas identified by prior study as undergoing protracted MIND maturation in typical development (Liang et al., Reference Liang, Sun, Xia, Zhao, Gong, Li, Liao, Cui, Duan, Pang, Wang, Yu, Bi, Chen, Chen, Chen, Chen, Cheng, Cheng and He2025). In contrast, PLS1− genes were mainly enriched in the cerebellum and thalamus during infancy and young adulthood, aligning with the late-maturing subcortical regions in the normative MIND trajectory (Liang et al., Reference Liang, Sun, Xia, Zhao, Gong, Li, Liao, Cui, Duan, Pang, Wang, Yu, Bi, Chen, Chen, Chen, Chen, Cheng, Cheng and He2025). These findings suggest that susceptibility windows for adolescent MDD may be brain-region-specific, with earlier vulnerability in cortical–limbic circuits driven by synaptic and neurodevelopmental transcriptional programs, and later vulnerability in subcortical regions governed by metabolic processes. Overall, by integrating macroscale MIND networks with microscale gene expression profiles, our study provides biological underpinnings associated with abnormal development trajectories of morphometric brain networks.
The application of regional MIND features within machine learning frameworks demonstrated promising potential for discriminating adolescents with MDD from HCs, contributing to the development of objective biomarkers for diagnosis. The optimal model achieved a high AUC of 0.824 and generalized well to the external independent dataset, demonstrating robust classification performance. Furthermore, the SHAP analysis provided transparent interpretability, highlighting the critical contributions of MIND in the prefrontal cortex and anterior cingulate cortex, which are extensively implicated in higher-order cognitive functions and emotional regulation (Tanji & Hoshi, Reference Tanji and Hoshi2008), in distinguishing adolescent MDD cases. These findings align with previous researches suggesting the involvement of these regions in the neurobiology of adolescent MDD (Li et al., Reference Li, Shen, Zhang, Li and Wu2025; Pannekoek et al., Reference Pannekoek, van der Werff, van den Bulk, van Lang, Rombouts, van Buchem, Vermeiren and van der Wee2014). Overall, this study underscores the potential utility of regional MIND metrics as biomarkers that could enhance early diagnosis, improve prognostic accuracy, and inform personalized intervention strategies for adolescent MDD.
Several limitations of our study should be mentioned. First, the sample size of this study was relatively small, which may affect the statistical power. Second, the classification performance, although promising, requires validation in larger, independent cohorts and diverse populations to ensure robustness and generalizability. Third, the transcriptional datasets from the AHBA database are limited to gene expression in the left hemisphere, and it is still unclear how adolescent MDD-related regional MIND changes are connected to gene signatures in the right hemisphere. Finally, the reliance on gene expression data from adult post-mortem sources might not fully capture adolescent-specific molecular dynamics; future work with adolescent-specific transcriptomic datasets would enhance biological relevance.
Conclusion
Using a novel MIND network model, this study identified abnormal cortical patterns in emotional and sensorimotor-related regions. The machine learning model based on MIND networks achieved strong performance in classification of adolescents with MDD versus controls. Furthermore, the connectome–transcriptome association analysis revealed that genes linked to MIND changes were enriched in neurobiologically relevant pathways and preferentially expressed in distinct cell types. Together, these findings advance our understanding of the neurobiological mechanism underlying adolescent MDD at both the macroscale network and the microscale molecular levels and highlight the potential of regional MIND measures as biomarkers for diagnosis.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S003329172610590X.
Data availability statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Funding statement
This study was supported by the National Natural Science Foundation of China (82502501) and the Sichuan Science and Technology Program (2025ZNSFSC1766).
Competing interests
The authors declare no competing financial interests.
Open access
