In this study, we aimed to investigate the shared molecular mechanisms underlying GMV changes in T2DM and MDD. We first conducted neuroimaging meta-analyses to identify GMV alterations specific to each disease and pinpoint the common changes occurring in both conditions. Following this, we utilized conjunctional false discovery rate (conjFDR) methods [35] to identify genes shared between the two diseases based on recent GWAS summary data for T2DM and MDD [28, 29]. Next, we performed transcriptome-neuroimaging association analyses with these shared genes, employing the AHBA to uncover the genetic drivers contributing to these structural brain changes. The identified shared genes were subsequently prioritized through expression-trait association analyses. Finally, knockout mouse models were employed to investigate the functional roles of the prioritized genes. The findings from this study not only enhance our understanding of the neurobiological mechanisms linking T2DM and MDD but also highlight potential genetic targets for future therapeutic interventions aimed at alleviating the cognitive and emotional impairments commonly associated with these diseases. An outline of the analytical framework is illustrated in Fig. 1.

A systematic search was conducted in the PubMed, Web of Science, and Embase databases to retrieve studies published prior to September 2024 using the following search terms: "voxel-based morphometry" or "VBM", and "gray matter volume" or "grey matter volume" or "GMV", and "major depressive disorder" or "major depression" or "MDD" or "type 2 diabetes mellitus" and "T2DM". To identify additional studies, we also manually searched the reference lists of the included articles. Studies were included if they met all of the following criteria: (1) published in peer-reviewed English journals; (2) employed the voxel-based morphometry (VBM) approach to analyze whole-brain GMV in patients with MDD or T2DM compared to healthy controls; (3) provided peak coordinates of GMV alterations along with the corresponding statistical values in Montreal Neurological Institute (MNI) or Talairach space, or reported null findings; and (4) included participants aged between 18 and 65 years. Studies were excluded if they met any of the following criteria: (1) non-original research, such as meta-analyses, editorials, or review articles; or (2) included participants with comorbidities involving other mental illnesses. In cases where patient groups overlapped with those of another study, the study with the larger sample size was retained. If an article reported multiple independent patient samples, these were treated as separate datasets. Our meta-analysis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [36], and the detailed study selection procedures are summarized in Fig. 2.

The quality of the included studies was evaluated using a 10-point checklist, which assessed several key factors, including the demographic and clinical characteristics of the subjects, the methods used for image acquisition and analysis, and the quality of the reported results and conclusions (Supplementary Table 1). Each criterion was scored as 1, 0.5, or 0, depending on whether it was fully, partially, or not met (Supplementary Tables 2, 3). Studies scoring above 9 were considered eligible and were included in subsequent analyses. Additionally, the following data were extracted from the included studies: demographic information (e.g., sample size, mean age, gender), clinical characteristics (e.g., illness duration, medication status), and methodological details (e.g., field strength, statistical threshold, scanning parameters). The peak coordinates and relevant statistics (e.g., t-values or their equivalents) were also recorded.

A T2DM dataset collected at Tianjin Medical University General Hospital (TMUGH) was included, comprising 156 right-handed individuals with T2DM and 148 healthy controls matched for age, sex, and handedness. Its inclusion in the meta-analysis increased the sample size and statistical power for detecting GMV alterations associated with T2DM. Detailed preprocessing and statistical procedures are provided in the Supplementary Materials.

A meta-analysis was performed using the Seed-based d Mapping with Permutation of Subject Images software (SDM-PSI, version 6.22; https://www.sdmproject.com/) to examine abnormal GMV changes in patients with T2DM and MDD. For each study, peak coordinates and corresponding statistical values (t-values, or P- and z-values converted to t-values via the SDM online converter) were utilized alongside the t-map from the TMUGH dataset to reconstruct full-brain effect-size images using default settings (full anisotropy = 1, full-width at half-maximum = 20 mm, voxel size = 2 mm, gray matter mask). Key steps included (1) generating multiple imputations for study and subject images, (2) performing group analyses on the imputed subject images within each study, and (3) conducting meta-analyses of the imputed study images using a random-effects model, with Rubin's rules applied to integrate the final meta-analysis results [37]. Age was included as a covariate to control for potential confounding effects across both patient and control groups in the meta-analysis, as T2DM patients are typically older than those with MDD. To control for family-wise error, a subject-based nonparametric permutation test (1000 permutations) was conducted with threshold-free cluster enhancement (TFCE), setting significance at FWE-TFCE P < 0.05 with a minimum cluster extent of 100 voxels. Finally, heterogeneity across studies was assessed using Cochran's Q test and I² statistics, and Egger's test was employed to evaluate publication bias in significant findings [38].

To uncover shared neurobiological mechanisms, we performed SDM-PSI's multimodal analysis [39] to identify common GMV changes across T2DM and MDD. Following the multimodal analysis, we conducted meta-regression analyses to examine how demographic and clinical factors moderated GMV changes within each condition. Specifically, for T2DM, we investigated the impact of variables such as the percentage of female participants, BMI, HbA1c levels, illness duration, and the proportion of patients on medication. For MDD, we assessed the percentage of female patients, onset age, illness duration, Hamilton Depression Rating Scale (HDRS) score, and the proportion of patients on medication. To assess the robustness of meta-analytic findings and explore potential sources of heterogeneity in MDD, we conducted subgroup analyses based on four consistently reported study-level variables: MRI field strength (1.5 vs. 3.0 T), scanner manufacturer (GE, Siemens, Philips), smoothing kernel size (8 mm vs. others), and medication status (medicated vs. unmedicated). These parameters were selected due to their frequent reporting and limited variability across included MDD datasets, enabling meaningful stratification. In contrast, subgroup analysis was not feasible for T2DM, given the smaller number of eligible studies (n = 15) and the limited availability of stratified metadata such as medication information.

GWAS summary statistics for T2DM were obtained from the DIAGRAM consortium, based on a large multi-ancestry meta-analysis involving 180,834 individuals with T2DM and 1,159,055 controls, partially derived from the UK Biobank [28]. For this study, we used the European ancestry subset, as it provided the largest sample size, comprising 80,154 cases and 853,816 controls. Depression summary statistics were obtained from a meta-analysis of six datasets, involving 371,184 individuals with depression and 978,703 controls, all of European ancestry [29]. To prevent sample overlap, which could introduce bias in the conjFDR analysis, participants from the UK Biobank were excluded from the depression dataset, and samples from 23andMe were removed due to access restrictions, resulting in a refined GWAS subset with 166,773 cases and 507,679 controls. Genome-wide genetic correlation between T2DM and MDD was estimated using linkage disequilibrium score regression (LDSC; https://github.com/bulik/ldsc). Analyses were restricted to HapMap3 SNPs, and LD scores were derived from the European reference panel of the 1000 Genomes Project.

Conditional FDR (condFDR) and conjFDR analyses were used to enhance discovery for each trait and identify shared genetic variants between T2DM and MDD, respectively. CondFDR, based on an empirical Bayesian framework, increases the power of GWAS by identifying significant loci and quantifying genetic overlap across complex traits. ConjFDR extends condFDR by prioritizing SNPs associated with a primary trait (e.g., T2DM) through conditional adjustment for a secondary trait (e.g., MDD), thereby leveraging the combined power of two GWAS to enhance variant identification [40]. For more detailed methodological descriptions, please refer to the Supplementary Materials. We set the significance threshold for condFDR analysis at 0.01 and conjFDR analyses at 0.05, following prior studies [41,42,43]. To minimize bias from complex linkage disequilibrium (LD) patterns, SNPs within the MHC (chr6: 25,119,106-33,854,733), 8p23.1 (chr8: 7,200,000-12,500,000), and MAPT (chr17: 40,000,000-47,000,000) were excluded before fitting the model [44]. To visualize the cross-trait genetic enrichment of T2DM and MDD, quantile-quantile (Q-Q) plots were generated at four increasingly P-value thresholds (P < 1, P < 0.1, P < 0.01, and P < 0.001), where leftward deflection from the expected line indicated cross-trait enrichment as the strength of association in the secondary trait increased.

To identify genes shared between T2DM and MDD, we used FUMA (https://fuma.ctglab.nl/) with the following steps. Independent significant SNPs were selected based on the conjFDR < 0.05 and were required to be independent of each other, with r < 0.6 [40, 45, 46]. Candidate SNPs were defined as those with conjFDR < 0.10 and LD r ≥ 0.6 with at least one independent significant SNP. These candidate SNPs were then mapped to protein-coding genes using three complementary mapping approaches: positional mapping, linking SNPs to genes within a 10 kb physical distance; eQTL mapping, associating SNPs with genes affected by allelic variation in expression; and chromatin interaction mapping, which uses 3D genome structure to link SNPs to both nearby and distant genes. Genes identified by any of these mapping methods were considered shared genes between T2DM and MDD.

To further investigate the genomic loci shared between T2DM and MDD and better understand the biological roles of candidate SNPs, we performed the following steps. First, we identified lead SNPs among the independent significant SNPs as those in LD with other significant SNPs at r < 0.1. For each lead SNP, a locus was established by including its associated independent SNPs and their related candidate SNP. Loci within 250 kb of each other were merged to form a single genomic locus, and LD information for this analysis was obtained from the European ancestry reference panel of the 1000 Genomes Project. Second, all candidate SNPs were annotated with combined annotation-dependent depletion (CADD) scores, RegulomeDB scores, and chromatin states. CADD scores assess the potential deleteriousness of each variant by integrating multiple annotations, such as evolutionary conservation and predicted functional impacts, with higher scores suggesting a greater likelihood of pathogenicity [47]. A CADD score threshold of 12.37 was applied to identify variants with potential pathogenic effects. RegulomeDB scores measure the regulatory potential of SNPs, with values ranging from 1-7; lower scores (e.g., 1a, 1b) represent stronger regulatory evidence, helping to pinpoint SNPs likely involved in gene regulation [48]. Chromatin states classify genomic regions based on epigenetic markers, identifying functional elements such as promoters, enhancers, and heterochromatic regions [49, 50].

Gene expression data were sourced from the AHBA database, a comprehensive repository that provides spatially mapped gene expression data across the human brain, derived from postmortem samples of six neurotypical adult donors [34] (Supplementary Table 4). This dataset encompasses over 20,000 genes, measured with approximately 58,000 probes across 3702 tissue samples from various brain regions. The abagen toolbox (https://www.github.com/netneurolab/abagen) [51] was used to extract and preprocess the gene expression data, following several key steps [52, 53]. First, gene probes were reannotated, and those that were unreliable or unmatched to genes were discarded [54]. An intensity-based filter was then applied, removing probes whose intensity was below background noise in at least 50% of tissue samples. The MNI coordinates of the tissue samples were aligned with the anatomical images using the alleninf package (https://github.com/chrisgorgo/alleninf). For each gene, the probe exhibiting the highest differential stability was selected as its representative. Gene expression values were subsequently normalized using a scaled robust sigmoid function, which was applied across genes for each sample and donor and across samples for each gene and donor. Additional normalization was performed within each distinct brain structure (cortex, subcortical/brain stem, and cerebellum) to adjust for differences in gene expression patterns among these regions. After preprocessing, the shared genes identified through conjFDR analysis for T2DM and MDD were extracted from the resulting gene expression matrix, and used for subsequent transcriptome-neuroimaging association analysis.

Using the unthresholded voxel-wise meta-analytic z-map representing GMV differences between patients with T2DM/MDD and healthy controls, we calculated the mean z-value within a 6 mm radius sphere centered at each tissue sample's MNI coordinate to represent the GMV difference for that sample. Within each disease group, cross-sample correlations between GMV differences and gene expression profiles were assessed using Pearson's correlation to identify genes whose expression levels correlated with GMV alterations in T2DM and MDD, respectively. To control for multiple comparisons, the Benjamini-Hochberg FDR method was applied (FDR P < 0.05). Genes associated with GMV alterations in both T2DM and MDD were identified by intersecting the results from the two separate cross-spatial correlation analyses, highlighting potential shared genetic factors underlying GMV changes in both conditions. To assess the significance of these shared genes, we conducted a permutation test (1000 iterations) using the Brain Surrogate Maps with Autocorrelated Spatial Heterogeneity (BrainSMASH) method [55], with significance set at P < 0.05.

To further refine the shared genes identified through the transcriptome-neuroimaging association analysis, we employed expression-trait association analyses to prioritize genes potentially relevant to GMV alterations in both T2DM and MDD. We used S-PrediXcan [56], a method that estimates genetically regulated gene expression in specific tissues and correlates it with complex traits, to conduct single-tissue expression-trait association analyses. This approach integrates GWAS summary data with pre-computed eQTL prediction models derived from individual brain tissues available in the Genotype-Tissue Expression (GTEx v8) dataset [57], including the amygdala, anterior cingulate cortex (BA24), caudate, cerebellar hemisphere, cerebellum, cortex, frontal cortex (BA9), hippocampus, hypothalamus, nucleus accumbens, putamen, cervical spinal cord C1, and substantia nigra. Next, we employed S-MultiXcan [58], an extension of S-PrediXcan, to combine results across tissues, increasing statistical power and identifying shared genetic effects across brain regions. Genes lacking significant brain eQTL associations were excluded, and significant expression-trait associations were identified using a Benjamini-Hochberg FDR-corrected threshold of P < 0.05.

To explore the functional relevance of shared genes between T2DM and MDD, we leveraged the Mouse Genome Informatics (MGI) database (https://www.informatics.jax.org/), a comprehensive international resource dedicated to laboratory mice. MGI database provides valuable information on gene mutations, knockout models, and associated phenotypic markers, enabling us to investigate the phenotypic consequences of gene loss in mice. Genes retrieved from in the MGI database were subsequently queried for human homologs using the Ensembl BioMart platform (https://mart.ensembl.org/) to ensure cross-species validity and translational relevance of our findings. Our analysis primarily focused on phenotypic changes observed in the nervous system and metabolic system, particularly those related to brain structure, function, or fasting circulating glucose level. By examining the abnormal behavioral and neurobiological phenotypes of the shared genes in knockout or mutant mouse model, we aimed to explore how the loss of these genes may affect behavior, such as activity levels, anxiety, or cognitive function, which could be relevant to both T2DM and MDD.