Introduction
Major depressive disorder (MDD) is a highly prevalent psychiatric disorder affecting more than 300 million individuals worldwide (Namiot et al., Reference Namiot, Smirnovová, Sokolov, Chubarev, Tarasov and Schiöth2024) and is projected to become a leading contributor to the global disease burden by 2030 (Malhi & Mann, Reference Malhi and Mann2018). It is characterized by core symptoms including depressed mood, anhedonia, and cognitive impairments (Gnanavel & Robert, Reference Gnanavel and Robert2013). Despite its substantial public health impact, the clinical presentation of MDD is highly heterogeneous (Lan et al., Reference Lan, Li, Zhu, Wu, Shen, Dai, Su and Si2025), and its etiology and underlying pathophysiological mechanisms remain poorly understood.
Accumulating evidence implicates immune activation and chronic inflammation in the pathophysiology of MDD (Bower & Kuhlman, Reference Bower and Kuhlman2023; Goldsmith, Bekhbat, Mehta, & Felger, Reference Goldsmith, Bekhbat, Mehta and Felger2023). Peripheral inflammatory disturbances, including elevated levels of cytokines such as interleukin-6 (IL-6), interleukin-2 (IL-2), and tumor necrosis factor-α (TNF-α), have been repeatedly reported, suggesting broad immune activation in affected individuals (Gędek et al., Reference Gędek, Modrzejewski, Materna, Iwański, Wichniak and Dominiak2025). These immune alterations may influence the brain through immune-to-central signaling pathways and blood–brain barrier (BBB) dysfunction, potentially contributing to structural abnormalities associated with MDD (Frodl & Amico, Reference Frodl and Amico2014; Gędek et al., Reference Gędek, Modrzejewski, Materna, Iwański, Wichniak and Dominiak2025; Jaeger et al., Reference Jaeger, Dohgu, Sultana, Lynch, Owen, Erickson, Shah, Price, Fleegal-Demotta, Butterfiled and Banks2009). Conversely, emerging neurobiological models propose that microglial activation may elevate central pro-inflammatory signaling, including TNF-α, thereby compromising BBB integrity and promoting broader neuroimmune dysregulation (Yang et al., Reference Yang, Ran, Li, Lin, Ma, Yang, Fu and Yang2022). Beyond classical pro-inflammatory cytokines, immune regulatory factors such as tissue inhibitor of metalloproteinases (TIMP; Sarkar et al., Reference Sarkar, Gharami, Mondal, Padmanabhan, Paidi, Srikumar and Biswas2025), growth-related factors including insulin-like growth factor-1 (IGF-1; Fernández-Pereira & Agís-Balboa, Reference Fernández-Pereira and Agís-Balboa2025), and angiogenic factors such as vascular endothelial growth factor (VEGF; J. Wang, Meng, Wang, & Li, Reference Wang, Meng, Wang and Li2025) have been shown to exert neuroprotective and neurotrophic effects. However, evidence linking immune markers to MDD remains inconsistent, and the complexity and interdependence of immune indicators make it challenging to clarify their causal roles, specific characteristics, and physiological significance in MDD (Jin et al., Reference Jin, Li, Jeong, Castro-Martinez and Zuker2024).
Structural and functional imaging studies over the past decade have substantially advanced our understanding of the neurobiological basis of MDD (Cui et al., Reference Cui, Li, Wang, Wu, Liu, Yu, Wang, Tang, Xia and Li2024; Zhu et al., Reference Zhu, Chen, Lu, Li, Wang, Cao, Chen, Chen, Chen, Chen, Cheng, Chu, Cui, Cui, Deng, Gong, Guo, He, Hu and Yu2024). Converging evidence supports the view that MDD is characterized by systemic biological dysregulation (Sforzini et al., Reference Sforzini, Marizzoni, Bottanelli, Kunšteková, Zonca, Saleri, Kose, Lombardo, Mariani, Nettis, Nikkheslat, Worrell, Zajkowska, Pointon, Cowen, Cavanagh, Harrison, Riva, Mondelli and Pariante2025), with prominent involvement of structural abnormalities in the central nervous system, particularly alterations in gray matter volume (GMV; Du, Tang, Gao et al., Reference Du, Tang, Gao, Li, Liu, Tang, Zhang, Lu, Zhang and Lui2025; Thomas-Odenthal et al., Reference Thomas-Odenthal, Ringwald, Teutenberg, Stein, Alexander, Bonnekoh, Brosch, Dohm, Flinkenflügel, Grotegerd, Hahn, Jansen, Leehr, Meinert, Pfarr, Renz, Schürmeyer, Stief, Straube and Kircher2024). Large-scale voxel-based morphometry studies have consistently reported regional GMV reductions in MDD, most notably in the prefrontal cortex, anterior cingulate cortex, and hippocampus (Brosch et al., Reference Brosch, Stein, Schmitt, Pfarr, Ringwald, Thomas-Odenthal, Meller, Steinsträter, Waltemate, Lemke, Meinert, Winter, Breuer, Thiel, Grotegerd, Hahn, Jansen, Dannlowski, Krug and Kircher2022; Lemke et al., Reference Lemke, Romankiewicz, Förster, Meinert, Waltemate, Fingas, Grotegerd, Redlich, Dohm, Leehr, Thiel, Enneking, Brosch, Meller, Ringwald, Schmitt, Stein, Steinsträter, Bauer and Dannlowski2022; Shen et al., Reference Shen, Wang, Chen, Lu, Li, Wang, Cao, Chen, Chen, Chen, Chen, Cheng, Chu, Cui, Cui, Deng, Gong, Guo, He and Yu2026; Yi et al., Reference Yi, Xia, Yi, Jia, Wei, Shen, Wu, Wang, Zhou, Li, Yan and Zhang2025) – regions critically involved in mood regulation, cognitive control, and stress processing (Zhao et al., Reference Zhao, Chen, Zhang, Xiao, Shah, Zhu, Yuan, Sun, Yue, Jia, Zhang, Kuang, Gong and Lui2017). Nevertheless, substantial inconsistencies persist across studies, limiting reproducibility and interpretability. These discrepancies are partly attributable to the inherent limitations of cross-sectional designs, which are unable to infer temporal dynamics or causal relationships (Savitz & Wellenius, Reference Savitz and Wellenius2023). Mendelian randomization (MR), which uses genetic variants as instrumental variables, offers a quasi-experimental framework for causal inference by reducing confounding and reverse causation under specific assumptions (Davey Smith & Hemani, Reference Davey Smith and Hemani2014; Davies, Holmes, & Davey Smith, Reference Davies, Holmes and Davey Smith2018; Du, Tang, Liu et al., Reference Du, Tang, Liu, Zhang and Lui2025). An increasing number of MR studies have investigated causal relationships between regional brain structure and MDD (Cai et al., Reference Cai, Li, Liu, Zheng, Zhong, Xue, Zhang, Zheng, Jin and Li2024; Zheng et al., Reference Zheng, Fang, Wang, Feng, Tang and Chen2025), as well as between inflammatory markers and MDD risk (X. Chen, Yao et al., Reference Chen, Yao, Cai, Fu, Li and Wu2022; Wang et al., Reference Wang, Jin, Cheng, Guan, Zheng and Zhang2023). However, existing work has largely focused on a limited subset of common immune markers, leaving the broader neuroimmune landscape insufficiently characterized.
Beyond regional abnormalities, MDD is increasingly conceptualized as a disorder of distributed brain networks rather than isolated brain regions (Gong & He, Reference Gong and He2015). Structural covariance networks (SCNs), which capture coordinated morphological variation across anatomically connected regions, provide a network-level perspective on brain organization (Chen et al., Reference Chen, Liu, Sun, Jiang, Yuan and Qing2025). Although SCN alterations have been reported in MDD, most studies rely on group-level analyses that obscure individual variability and preclude direct associations with clinical or biological markers (Long et al., Reference Long, Qin, Pan, Fan and Li2024). To address this limitation, individualized structural covariance networks (IDSCNs) have been proposed and systematically validated (Han et al., Reference Han, Xue, Chen, Xu, Li, Song, Guo, Fang, Zheng, Zhou, Chen, Wei, Zhang and Cheng2023; Long et al., Reference Long, Qin, Pan, Fan and Li2024; Luo et al., Reference Luo, Hu, Qiu, Li, Wang, Lan, Mai, Chen, Liu, Zhang, Chen, You, Zeng, Liang, Chen, Lu, Zhou and Ning2025; Qin et al., Reference Qin, Li, Zhang, Yin, Wu, Pan, Chen, Roberts, Sweeney, Huang, Gong and Jia2024). Prior studies have demonstrated disrupted network centrality and connectivity across multiple large-scale systems in MDD (Long et al., Reference Long, Qin, Pan, Fan and Li2024), yet these investigations have primarily focused on associations with symptom severity (Yang et al., Reference Yang, Kumar, Nickerson, Du, Wang, Chen, Li, Pizzagalli and Ma2021), while the relationships between individualized network architecture and objective peripheral biomarkers remain poorly understood. In particular, it remains unclear whether this structure-driven subtype heterogeneity corresponds to different inflammatory immune characteristics and differences in response to antidepressants. Further, although peripheral immune markers have shown potential clinical relevance in MDD, including associations with treatment response (Xu et al., Reference Xu, Liang, Sun, Zhang, Shan, Ge and Xia2023), their explanatory value remains limited when considered in isolation. Moreover, while distinct immune-inflammatory profiles have been reported across neuroimaging-defined MDD subgroups (Colombo et al., Reference Colombo, Calesella, Bravi, Fortaner-Uyà, Monopoli, Tassi, Carminati, Zanardi, Bollettini, Poletti, Lorenzi, Spadini, Brambilla, Serretti, Maggioni, Fabbri, Benedetti and Vai2024), the organization of these alterations into biologically meaningful immune-inflammatory axes and their relevance to subtype heterogeneity and treatment outcomes remains poorly understood. Interpretable machine-learning approaches, such as extreme gradient boosting (XGBoost), offer a powerful framework for quantifying the relative contributions of interacting immune-inflammatory systems and may facilitate a more mechanistic understanding of MDD heterogeneity (Xie et al., Reference Xie, Che, Chen, Mao and Chen2025).
Here, we combined MR, structural neuroimaging, and longitudinal clinical follow-up to investigate the relationships among peripheral immune-inflammatory systems, brain structural networks, and MDD. We further examined whether neuroimaging-defined subtypes exhibit distinct immune-inflammatory characteristics and treatment trajectories. By linking immune-inflammatory axes to brain structural heterogeneity and clinical outcomes, this study aims to advance a biologically grounded framework for patient stratification in MDD.
Methods
This study employed an integrative design combining MR, multimodal neuroimaging, and a prospective follow-up cohort to investigate the relationships among inflammatory and immune-related markers, brain structure, and MDD. Bidirectional two-sample MR analyses were first conducted using large-scale genome-wide association study (GWAS) summary statistics. A clinical neuroimaging cohort was then used to examine brain structural alterations and immune-inflammatory profiles and to identify MDD subtypes. Finally, machine learning and survival analyses were applied to evaluate the predictive and prognostic value of these multimodal features. The study workflow is illustrated in Figure 1, with detailed methods provided in the Supplementary Methods.
The workflow of this study. Note: SNPs, ‘single nucleotide polymorphisms’; GMV, ‘gray matter volume’; MR, ‘Mendelian randomization’; IVW, ‘inverse variance weighting’; HCs, ‘healthy controls’; MDD, ‘major depressive disorder’; IDSCN, ‘individualized structural covariance network’.
Figure 1. Long description
The flowchart is divided into two main horizontal sections.
Top Section: MR analysis.
On the left, Instrumental variable SNP s are shown with three assumptions. Assume 1 points directly to Exposure, which includes icons for MD D, GM V, and blood markers. Assume 2 and Assume 3 are marked with red crosses, indicating blocked paths to Confounders and Outcome respectively. A bidirectional arrow connects Exposure and Outcome, representing Bidirectional MR. Confounders at the top center point to both Exposure and Outcome. Below this, two text boxes list MR analysis methods including IV W, MR egger, and weighted modes, and explain that pairwise two-sample MR analyses were conducted among the 3 phenotypes.
Bottom Section: Clinical and Biological Workflow.
Starting on the left, participants are categorized into HCs and MDD groups.
The top path shows MRI of whole brain, followed by Preprocess, then Construction of IDSCN. This involves mapping brain regions to create rSCN and pSCN matrices to calculate z-scores for individualized structural covariance networks.
The bottom path shows Collection of blood samples, Centrifugation, and Quantitative testing for cytokines and chemokines such as IL-1 beta, INF-gamma, and TNF-alpha.
Both paths converge on the right for IDSCN analysis, followed by Machine learning and survival analysis, visualized with a bar chart and a Kaplan-Meier survival curve.
MR analysis
Data sources
To investigate genetically determined and potentially causal associations among inflammatory and immune-related markers, brain structure, and MDD, we conducted MR analyses using large-scale GWAS summary statistics. Summary data for MDD were obtained from the most recent and largest meta-analysis, including 688,808 cases and 4,364,225 controls (Major Depressive Disorder Working Group of the Psychiatric Genomics Consortium, 2025). GWAS summary statistics for GMV across 110 brain regions were derived from the UK Biobank imaging genetics study reported by Wang et al. (Reference Wang, Martins-Bach, Alfaro-Almagro, Douaud, Klein, Llera, Fiscone, Bowtell, Elliott, Smith, Tendler and Miller2022), comprising 35,273 individuals of European ancestry. In addition, GWAS datasets for inflammatory and immune-related markers were retrieved from publicly available repositories. The detailed information regarding the sample sources and sizes of the GWAS database is presented in Supplementary Table S1.
Instrumental variables
Single nucleotide polymorphisms (SNPs) associated with each exposure were selected as instrumental variables based on genome-wide significance thresholds. Linkage disequilibrium clumping was performed using the 1,000 Genomes Project reference panel to ensure the independence of selected instruments. Instrument strength was assessed using F-statistics, and weak instruments were excluded (Cheng et al., Reference Cheng, Yang, Shi, Yeung, Yang, Peng and Liu2020). SNPs potentially associated with confounding traits were removed using PhenoScanner v2 (Kamat et al., Reference Kamat, Blackshaw, Young, Surendran, Burgess, Danesh, Butterworth and Staley2019).
MR statistical analysis
Two-sample MR analyses were conducted to evaluate potential causal relationships among inflammatory and immune-related markers, regional GMV, and MDD, based on the core MR assumptions (Figure 1). Inverse variance weighting (IVW) was used as the primary analytical method, complemented by MR-Egger, weighted median, weighted mode, and simple mode approaches to assess robustness (Yao et al., Reference Yao, Zhang, Dong, Wang, Zhang, Guo, Guo and Yang2022; Zhao, Zhao, Cao, & Tu, Reference Zhao, Zhao, Cao and Tu2023; Zhu et al., Reference Zhu, Sun, Guo, Zhou, Jiang, Shen, Zhou, du and Zhou2023). Heterogeneity and horizontal pleiotropy were evaluated using Cochran’s Q test, the MR-Egger intercept, and MR-PRESSO (Burgess & Thompson, Reference Burgess and Thompson2017; Yao et al., Reference Yao, Zhang, Dong, Wang, Zhang, Guo, Guo and Yang2022). Random-effects IVW models were applied when significant heterogeneity was detected, and false discovery rate (FDR) correction was used to account for multiple testing (Zhou et al., Reference Zhou, Gao, Zhang, Xu, Wang, Wang, Tan, Luo, Zhou, Shuai, Cai, Zheng, Wang, Duan and Wu2025).
Participants and clinical assessment
Participants
A total of 122 patients with MDD, aged 18–65 years, were recruited from the Mental Health Center, West China Hospital of Sichuan University between May 2021 and August 2024. MDD was diagnosed using the Mini International Neuropsychiatric Interview (M.I.N.I.) and confirmed by two experienced psychiatrists (Roehr, Reference Roehr2013). Sixty-six age- and sex-matched HCs were recruited from the local community and screened using the M.I.N.I. to exclude lifetime psychiatric disorders (Sheehan et al., Reference Sheehan, Lecrubier, Sheehan, Amorim, Janavs, Weiller, Hergueta, Baker and Dunbar1998). All participants were right-handed and had not received psychological treatment or antidepressant medication for at least 2 weeks before MRI scanning, clinical assessment, and blood sampling.
To evaluate the generalizability of our findings, an independent external validation cohort was additionally included from the REST-meta-MDD project data of the Depression Imaging REsearch ConsorTium (DIRECT), a multi-center dataset collected across China (X. Chen, Lu et al., Reference Chen, Lu, Li, Li, Wang, Castellanos, Cao, Chen, Chen, Cheng, Cui, Deng, Fang, Gong, Guo, Hu, Kuang, Li, Li and Yan2022; Yan et al., Reference Yan, Chen, Li, Castellanos, Bai, Bo, Cao, Chen, Chen, Chen, Cheng, Cheng, Cui, Duan, Fang, Gong, Guo, Hou, Hu and Zang2019). We selected a subsample of first-episode, medication-naïve patients with MDD and corresponding HCs for the present analysis. After applying inclusion criteria, including availability of complete imaging and clinical data, right-handedness, and absence of prior psychiatric treatment, a total of 295 MDD patients and 504 age- and sex-matched HCs were included in the final analysis. Detailed demographic and clinical characteristics of the validation sample are presented in Supplementary Table S2.
The study was approved by the Ethics Committee of West China Hospital, Sichuan University (No. #2020:1054), and written informed consent was obtained from all participants or their legal guardians. Detailed inclusion and exclusion criteria are provided in the Supplementary Methods.
Clinical and demographic assessment
Two trained raters collected demographic data and administered clinical scale assessments to all participants before the MRI scan. All participants were assessed using the 17-item Hamilton Depression Scale (HAMD-17), the 14-item Hamilton Anxiety Scale (HAMA-14), the 9-item Patient Health Questionnaire (PHQ-9), and the Snaith–Hamilton Pleasure Scale (SHAPS). Subsequently, peripheral blood samples were collected by an experienced laboratory medicine technologist.
Following the baseline assessments, patients with MDD received antidepressant treatment as determined by their treating clinicians in routine clinical practice (Supplementary Table S3). Given that antidepressant medications typically require at least 1 month to exert clinically meaningful effects, treatment response was monitored through monthly follow-up evaluations using the PHQ-9 for up to 3 months. Treatment response and improvement were defined as a reduction of at least 5 points in PHQ-9 scores relative to baseline, consistent with a clinically meaningful improvement in depressive symptoms. Accordingly, a patient was considered a responder when this threshold was reached at any follow-up time point during the observation period (Gliklich, Leavy, & Li, Reference Gliklich, Leavy and Li2020).
Event/censoring data
Participants who did not experience remission during the 3-month follow-up were treated as right-censored at their last available assessment. Participants who were lost to follow-up before experiencing remission were censored at the time of their last completed assessment. Participants who completed the study without remission were censored at the end of the 3-month follow-up (Supplementary Table S4).
Inflammatory and immune-related marker measurement
Peripheral blood was collected from all participants, and serum levels of inflammatory and immune-related markers were measured using a multiplex bead-based antibody array (RayPlex; https://www.raybiotech.com/). Markers were organized into predefined biological functional categories for downstream analysis, based on their established roles in immune and inflammatory regulation (Murphy & Weaver, Reference Murphy and Weaver2016). Core inflammatory axis: granulocyte-macrophage colony-stimulating factor (GM-CSF), interferon-γ (IFN-γ), interleukin (IL)-6, C-reactive protein (CRP), tumor necrosis factor-α (TNF-α), TNF-β, and macrophage migration inhibitory factor (MIF). Adaptive immunity axis: transforming growth factor-β1 (TGF-β1), IL-2, IL-4, IL-5, IL-10, IL-12p70, IL-13, IL-17, IL-21, IL-23, and IL-28A. Chemokine axis: IL-8, monocyte chemoattractant protein-1 (MCP-1), macrophage inflammatory protein-1β (MIP-1β), MIP-3α, regulated upon activation normal T cell expressed and secreted (RANTES), interferon-γ–inducible protein 10 (IP-10), and pulmonary and activation-regulated chemokine (PARC). Tissue repair axis: IL-11, IL-22, insulin-like growth factor-1 (IGF-1), angiopoietin-1, angiopoietin-2, and VEGF. Fibrosis axis: plasminogen activator inhibitor-1 (PAI-1), tissue inhibitors of metalloproteinases-1 (TIMP-1), and TIMP-2. Other: IL-1β, IL-1R4, 4-1BB, agouti-related peptide (AgRP), Fas, and leptin. To improve distributional normality and reduce the influence of skewness, all biomarker concentrations were transformed using the natural logarithm before statistical analyses (Sharp, Reference Sharp2024; Zhang et al., Reference Zhang, Cao, Mu, Liu, Gao, Han, Zhai, Zhou, Ni, Yao, Li, Jin, Zhang, Cui, Shen and Zhu2022).
MRI acquisition
MRI scans were performed on a 3.0 Tesla whole-body scanner (Siemens Prisma, Erlangen, Germany) with a 64-channel head coil. High-resolution 3D-T1-weighted images were acquired for all participants. Images were visually inspected to exclude artifacts, and subsequent analyses focused on GMV differences between MDD patients and HCs.
IDSCN analysis
The comparison of demographic and clinical scale data was analyzed using R (version 4.4.2). IDSCNs were constructed using a network-template-perturbation framework based on the Automated Anatomical Labeling atlas comprising 90 brain regions (Tzourio-Mazoyer et al., Reference Tzourio-Mazoyer, Landeau, Papathanassiou, Crivello, Etard, Delcroix, Mazoyer and Joliot2002). Patient-specific deviations from the healthy control reference network were quantified as Z-scores for each structural covariance edge. To identify disorder-related alterations, edge-wise comparisons between patients with MDD and healthy controls were performed, and significantly altered edges were retained after Bonferroni correction (p < 0.05). These abnormal edges were subsequently mapped onto the seven large-scale functional networks defined by the Yeo-7 atlas to characterize their network-level distribution (Yeo et al., Reference Yeo, Krienen, Sepulcre, Sabuncu, Lashkari, Hollinshead, Roffman, Smoller, Zöllei, Polimeni, Fischl, Liu and Buckner2011). MDD subtypes were identified using K-means clustering based on the profiles of significantly altered IDSCN edges. To determine the optimal number of clusters, K-means clustering was performed iteratively with the number of clusters ranging from 2 to 10 (Hu, Zhu, & Zhang, Reference Hu, Zhu and Zhang2025). The clustering solution with the highest silhouette coefficient was selected as the optimal model, thereby providing a data-driven estimate of subtype number. Patients were assigned to subtypes according to this optimal clustering solution. To evaluate the reproducibility of the identified subtypes, the same clustering procedure was applied to an independent external validation cohort derived from the DIRECT dataset, in which regional GMV measures were harmonized across sites using the ComBat algorithm to mitigate scanner- and site-related effects before IDSCN construction and subtype classification (Fortin et al., Reference Fortin, Cullen, Sheline, Taylor, Aselcioglu, Cook, Adams, Cooper, Fava, McGrath, McInnis, Phillips, Trivedi, Weissman and Shinohara2018). Clinical differences between subtypes were assessed using independent-samples t-tests, and statistical significance was defined as p < 0.05.
Immune-inflammatory axis importance analysis
Given the extensive interactions and interdependencies among immune-inflammatory pathways, analyses based solely on individual biomarkers may not adequately capture the coordinated biological processes underlying MDD heterogeneity. To characterize the relative contributions of predefined immune systems to subtype differentiation, inflammatory markers were first grouped into five biologically informed axes. Specifically, for each axis, a composite score was calculated as the average of the constituent markers and subsequently standardized (z-transformed) prior to statistical analyses. This strategy follows the general approach of previous studies that summarized biologically related molecular features into standardized composite scores to represent coordinated pathway activity (Moisan et al., Reference Moisan, Foury, Dexpert, Cole, Beau, Forestier, Ledaguenel, Magne and Capuron2021).
XGBoost model was subsequently implemented to evaluate the importance of these inflammatory axes in distinguishing the identified MDD subtypes. XGBoost is an ensemble learning algorithm based on gradient-boosted decision trees that can effectively model complex nonlinear relationships and interactions among predictors while maintaining strong predictive performance (Xie et al., Reference Xie, Che, Chen, Mao and Chen2025). To enhance model interpretability, Shapley Additive exPlanations (SHAP) values were calculated, and the mean absolute SHAP value of each inflammatory axis was used to quantify its relative contribution to subtype classification (https://metricgate.com/docs/shap-tree-explainer/). Axes with larger SHAP values were considered more important in driving subtype differentiation.
Time-to-event analysis
To further investigate whether the identified subtypes differed in antidepressant treatment outcomes among MDD patients who participated in the follow-up, Kaplan–Meier survival analysis was performed to examine time to response. Survival curves were generated to visualize cumulative response probabilities over the follow-up period, and differences between subtypes were assessed using the log-rank test.
Results
Demographic and clinical data
The demographic information and clinical data of all participants are shown in Table 1. Fifty-nine MDD patients met the response criteria during the study period, and the remainder were considered censored samples. At baseline, there were no statistically significant differences in age, sex, or TIV between the MDD and HC groups. The MDD group showed fewer years of education but higher scores on HAMD-17, HAMA-14, PHQ-9, and SHAPS with the HC group. Comparisons of ln-transformed biomarker levels revealed a distinct immune-inflammatory profile in patients with MDD relative to HCs. Specifically, levels of IFN-γ and TNF-β within the core inflammatory axis, together with the adaptive immunity-related cytokines IL-4 and IL-5, were significantly elevated in the MDD group. Consistent with this pattern, chemokine levels, including MIP-1β and MIP-3α, were also increased in MDD (all p < 0.05, Bonferroni corrected). Furthermore, MDD was characterized by higher levels of the angiogenic and tissue-remodeling factors angiopoietin-1 and VEGF. In contrast, levels of TIMP-1 and IL-1R4 were significantly lower in patients with MDD than in HCs (all p < 0.05, Bonferroni corrected). No other markers remained significant after Bonferroni correction.
Baseline demographic and clinical characteristics and levels of inflammatory and immune-related markers of the participants
Table 1. Long description
The table consists of five columns: Variables, MDD (n = 122), HCs (n = 66), t/chi-squared, and p-Values.
Demographic and Clinical Variables:
* Age (years): MDD 28.78 plus or minus 8.56, HC s 27.65 plus or minus 8.86, p = 0.401.
* Sex (male/female): MDD 37/85, HCs 22/44, p = 0.676.
* TIV (mm super 3): MDD 1,452.07 plus or minus 126.14, HCs 1,486.45 plus or minus 149.87, p = 0.116.
* Education (years): MDD 15.60 plus or minus 2.42, HCs 16.39 plus or minus 1.97, p = 0.015.
* Clinical scales (H AMD-17, HAMA-14, PHQ-9, SHAPS) all show highly significant differences (p < 0.001) with higher scores in the MDD group.
Inflammatory and Immune Markers (selected significant results p < 0.05 after Bonferroni correction):
* Core inflammatory axis: IFN-gamma (M DD 2.66 vs HCs 2.31) and TNF-beta (M DD 6.44 vs HCs 5.77) are significantly higher in MD D.
* Adaptive immunity axis: IL-4 (M DD 5.11 vs HCs 4.72) and IL-5 (M DD 3.40 vs HC s 2.25) are significantly higher in M D D.
* Chemokine axis: MIP-1 beta (M DD 6.54 vs HCs 5.77) and MIP-3 alpha (M DD 5.95 vs HCs 5.53) are significantly higher in MD D.
* Tissue repair axis: Angiopoietin-1 (M DD 9.57 vs HCs 8.92) and VEGF (M DD 5.87 vs HCs 5.63) are significantly higher in MD D.
* Fibrosis axis: TIMP-1 is significantly lower in MDD (11.37) compared to HCs (11.8).
* Other: IL-1 R4 is significantly lower in MDD (2.67) compared to HCs (3.11).
Note: MDD, major depressive disorder; HCs, healthy controls; TIV, total intracranial volume; HAMD-17, Hamilton Depression Rating Scale, 17 items; HAMA-14, Hamilton Anxiety Rating Scale, 14 items; PHQ-9, Patient Health Questionnaire-9; SHAPS, Snaith-Hamilton Pleasure Scale; TNF, tumor necrosis factor; IL, interleukin; IFN-γ, interferon-gamma; CRP, C-reactive protein; MIF, macrophage migration inhibitory factor; GM-CSF, granulocyte-macrophage colony-stimulating factor; TGF-β, transforming growth factor beta; MCP-1, monocyte chemoattractant protein-1; MIP, macrophage inflammatory protein; RANTES, regulated upon activation normal T cell expressed and secreted; IP-10, interferon-gamma-induced protein 10; PARC, pulmonary and activation-regulated chemokine; IGF, insulin-like growth factor; VEGF, vascular endothelial growth factor; PAI-1, plasminogen activator inhibitor-1; TIMP, tissue inhibitor of metalloproteinases; 4-1BB, tumor necrosis factor receptor superfamily member 9; AgRP, agouti-related peptide; Fas, Fas (TNF receptor superfamily member 6).
* p < 0.05 after Bonferroni correction.
MR analysis
Bidirectional MR analyses were conducted to examine causal relationships among peripheral immune markers, regional GMV, and MDD. Instrumental variables were sufficiently strong (F-statistics >10; Supplementary Tables S5–S10).
Genetically predicted MDD was associated with altered circulating IGF-1 levels (ORIVW = 0.95, p FDR = 0.013; Figure 2a and Supplementary Figure S1). Structural variations in key cortico-striatal and limbic regions were causally linked to MDD risk (p FDR < 0.05; Figure 2a and Supplementary Figures S2–S6). Forward MR showed that certain inflammatory and immune-related markers, such as IL-5 and leptin, were negatively associated with GMV alterations in the right cingulate gyrus, posterior division, and bilateral putamen, respectively (p FDR < 0.05; Figure 2, Supplementary Figures S7–S9, and Supplementary Tables S11–S12). However, there were no significant causal associations between GMV alterations and the levels of immune profiles.
Causal links among inflammatory and immune-related markers, cortical structure, and MDD. MDD is associated with a decrease in IGF-1 levels, and changes in GMV in certain brain regions contribute to the risk of MDD. The changes in some inflammatory and immune-related marker levels can lead to specific GMV variations in different brain regions. Note: MDD, ‘major depressive disorder’; IVW, ‘inverse variance weighting’; OR, ‘odds ratio’; FDR, ‘false discovery rate’; IGF, ‘insulin-like growth factor’; MIP, ‘macrophage inflammatory protein’; IL, ‘interleukin’; IP-10, ‘interferon-gamma-induced protein 10’; PAI-1, ‘plasminogen activator inhibitor-1’; RANTES, ‘regulated upon activation normal T cell expressed and secreted’.
Figure 2. Long description
A forest plot table with nine rows and seven columns: Exposure, Outcome, Method, nsnp, a graphical plot, OR (95% CI), pval, and FDR. P. The x-axis of the plot ranges from 0.8 to 1.2, with a dashed vertical line at 1.0. Values to the left are labeled protective factor and values to the right are risk factor.
* Row 1: Exposure MD D, Outcome IGF-1, Method IV W, nsnp 171. Plot point is left of 1.0. OR 0.95, pval 6.62E-04, FDR. P 0.013.
* Row 2: Exposure Right middle frontal gyrus, Outcome MD D, Method IV W, nsnp 24. Plot point is right of 1.0. OR 1.08, pval 1.78E-03, FDR. P 0.047.
* Row 3: Exposure Left subcallosal cortex, Outcome MD D, Method IV W, nsnp 44. Plot point is right of 1.0. OR 1.06, pval 7.10E-04, FDR. P 0.045.
* Row 4: Exposure Right putamen, Outcome MD D, Method IV W, nsnp 46. Plot point is left of 1.0. OR 0.96, pval 2.14E-03, FDR. P 0.047.
* Row 5: Exposure Left ventral striatum, Outcome MD D, Method IV W, nsnp 55. Plot point is left of 1.0. OR 0.95, pval 8.19E-04, FDR. P 0.045.
* Row 6: Exposure Right ventral striatum, Outcome MD D, Method IV W, nsnp 43. Plot point is left of 1.0. OR 0.96, pval 1.96E-03, FDR. P 0.047.
* Row 7: Exposure IL-5, Outcome Right cingulate gyrus posterior division, Method IV W, nsnp 4. Plot point is left of 1.0. OR 0.94, pval 3.07E-05, FDR. P 0.010.
* Row 8: Exposure Leptin, Outcome Left putamen, Method IV W, nsnp 11. Plot point is far left with an arrow. OR 0.78, pval 2.73E-04, FDR. P 0.045.
* Row 9: Exposure Leptin, Outcome Right putamen, Method IV W, nsnp 11. Plot point is far left with an arrow. OR 0.78, pval 1.14E-04, FDR. P 0.038.
Sensitivity analyses revealed no evidence of horizontal pleiotropy (all p > 0.05; Supplementary Table S13).
IDSCN analysis
IDSCN analysis identified 37 edges showing significant differences between the MDD group and the HC group after Bonferroni correction (p < 0.05). Among these edges, those involving the left caudate nucleus accounted for the largest proportion (Figure 3a,b). Mapping these connections onto the Yeo-7 functional network atlas revealed that default mode network (DMN) to limbic network connections and limbic network self-connections were the most dominant alterations (8 edges; Figure 3c). K-means clustering based on the profile of abnormal edges further partitioned the MDD patients into two distinct subtypes: one characterized by predominantly decreased connectivity (subtype 1, hypoconnectivity subtype; n = 49) and the other by predominantly increased connectivity (subtype 2, hyperconnectivity subtype; n = 73; Figure 3d). The robustness of this two-subtype classification was further supported by its replication in an independent, larger cohort of first-episode, drug-naïve MDD patients from the DIRECT dataset (Supplementary Figures S10–S11). Notably, the hypoconnectivity subtype exhibited significantly higher scores on both the HAMA-14 and SHAPS compared with the hyperconnectivity subtype (both p < 0.05; Figure 3e,f).
Significant differences in individual structural covariance networks distinguished MDD subtypes. (a–c) Spatial distribution of significant IDSCNs and their mapping to Yeo’s 7 networks. (d) Group-level differences in connection-wise Z-scores between two subtypes in MDD. (e–f) Comparison of HAMA-14 and SHAPS scores between subtype 1 and subtype 2, with subtype 1 showing significantly higher scores than subtype 2 for both measures. Note: L, ‘left’; R, ‘right’; HAMA-14, ‘Hamilton Anxiety Rating Scale’, 14 items; SHAPS, ‘Snaith–Hamilton Pleasure Scale’.
Figure 3. Long description
A multi-panel figure consisting of six sections labeled A through F.
Panel A displays eight anatomical brain models from various perspectives (lateral, medial, dorsal, and ventral) showing blue nodes and connecting gold lines representing structural covariance networks. Labels include L for left and R for right.
Panel B is a radial chord diagram. The outer ring is divided into colored segments representing Yeo’s 7 networks: Visual (pink), Somatomotor (light blue), Dorsal Attention (green), Ventral Attention (teal), Limbic (purple), Frontoparietal (orange), and Default (dark blue). Internal ribbons connect specific brain regions such as CAU . L, AMYG. R, and PHG. L.
Panel C is a triangular heat map matrix showing the count of connections between the 7 networks. The highest counts are 8 for Limbic-Limbic and Limbic-Default connections.
Panel D is a line graph showing Mean Z-score on the Y-axis (ranging from minus 2 to 2) across various brain region pairs on the X-axis. Subtype 1 (blue line) remains consistently lower, near minus 0.5, while Subtype 2 (red line) is higher, fluctuating between 0 and 2. Shaded areas represent error margins.
Panel E and F are raincloud plots comparing HAMA minus 14 and SHAPS scores. In both plots, Subtype 1 (blue) shows higher median scores and distribution density compared to Subtype 2 (red). Panel E shows a p-value of 0.047 and Panel F shows a p-value of 0.045, indicating statistical significance.
Inflammatory and immune profiles and treatment response differ across MDD subtypes
XGBoost was then applied to evaluate the contribution of five predefined inflammatory and immunity-related axes. The core inflammatory axis exhibited the highest mean SHAP value (0.356), followed by the tissue repair axis (0.289), chemokine axis (0.286), adaptive immunity axis (0.206), and fibrosis axis (0.096; Figure 4a,b). The mean scores of the core inflammatory, chemokine, adaptive immune, and tissue repair axes were consistently elevated in subtype 2, whereas they were reduced in subtype 1. This trend was the exact opposite on the fibrosis axis.
Differential Immune-inflammatory axis contributions and treatment response among these two types in MDD. (a) Bar plot of mean absolute SHAP values for each immune-inflammatory axis. The core inflammatory axis ranks highest, indicating its primary role in distinguishing the two MDD subtypes. Values represent the average impact of each axis on model output across all samples. (b) SHAP beeswarm plot showing the distribution of SHAP values for each axis. (c) Comparison of the composite inflammatory score (mean score) between the two MDD subtypes. (d) Kaplan–Meier curves for time to antidepressant response. Subtype 1 showed a faster and higher cumulative response probability compared to subtype 2.
Figure 4. Long description
Panel A is a horizontal bar plot showing mean absolute SHAP values. From top to bottom, the axes are: Core inflammatory axis at 0.356, Tissue repair axis at 0.289, Chemokine axis at 0.286, Adaptive immunity axis at 0.206, and Fibrosis axis at 0.096.
Panel B is a beeswarm plot of SHAP values ranging from negative 0.5 to 0.75. A color scale on the right indicates feature value from low purple to high yellow. The distribution shows the impact of each axis on the model output for individual samples.
Panel C is a grouped bar chart comparing Mean Score across five axes for Subtype 1 in light blue and Subtype 2 in pink. Subtype 2 shows positive scores for Adaptive immunity, Chemokine, Core inflammatory, and Tissue repair axes, while Subtype 1 shows negative scores for these same axes. The Fibrosis axis shows the opposite trend with a high positive score for Subtype 1 and a negative score for Subtype 2.
Panel D is a Kaplan-Meier survival curve showing Cumulative response rate percentage over Time in months from 0.0 to 3.0. Subtype 1 in blue maintains a higher cumulative response rate over time compared to Subtype 2 in pink, with both curves showing a step-down pattern at 0.5 month intervals.
Importantly, the two subtypes demonstrated significantly different time-to-response trajectories. Subtype 2 showed a faster attainment of clinical response over time, with a higher cumulative probability of response compared with subtype 1 throughout follow-up (χ2 = 4.605, p = 0.032; Figure 4d).
Discussion
This study provides an integrative framework linking peripheral immune-inflammatory systems, brain structural networks, and clinical heterogeneity in MDD. Using bidirectional MR analyses, we identified genetically informed associations between circulating immune markers and brain regions implicated in cortico-striatal, limbic, and large-scale intrinsic networks, supporting a reciprocal relationship between immune signaling and brain structure. Building on these findings, IDSCN analysis revealed two reproducible neuroimaging-defined MDD subtypes characterized by distinct connectivity patterns. Importantly, these subtypes exhibited divergent immune-inflammatory axis profiles and treatment-response trajectories, with the core inflammatory axis showing the strongest contribution to subtype differentiation. Collectively, our findings suggest that immune-inflammatory dysregulation and brain structural network alterations jointly contribute to the biological and clinical heterogeneity of MDD, providing a potential framework for biologically informed patient stratification.
Causal links between immune signaling, cortical structure, and MDD vulnerability
In this study, bidirectional MR provided genetically informed evidence for causal interactions among MDD, peripheral inflammatory and immune-related markers, and brain gray matter architecture. Notably, our findings suggest that genetically predicted MDD is associated with reduced circulating IGF-1, despite previous observational reports of elevated IGF-1 levels in MDD (Aguiar-Oliveira, Boguszewski, Rovaris, & Donato, Reference Aguiar-Oliveira, Boguszewski, Rovaris and Donato2026). Population-based studies have indicated a U-shaped relationship between IGF-1 and depressive risk, with both low and high levels conferring vulnerability in different subgroups (Chigogora et al., Reference Chigogora, Zaninotto, Kivimaki, Steptoe and Batty2016; Sievers et al., Reference Sievers, Auer, Klotsche, Athanasoulia, Schneider, Nauck, Völzke, John, Schulz, Freyberger, Friedrich, Biffar, Stalla, Wallaschofski and Grabe2014), likely reflecting heterogeneity related to age, sex, illness stage, and treatment exposure. In this context, our MR results support the interpretation that IGF-1 reduction may characterize early or medication-naïve stages of MDD. In contrast, elevated IGF-1 observed in clinical samples may reflect compensatory or treatment-related processes.
Consistent with prior neuroimaging literature, genetically influenced variations in GMV within the middle frontal gyrus, subcallosal cortex, putamen, and ventral striatum were causally associated with MDD risk (Liu et al., Reference Liu, Sun, Yang, Li, Gao, Gong, Zhang and Lui2024). Importantly, reverse MR analyses did not support a causal effect of MDD on GMV, indicating that these structural alterations likely represent pre-existing vulnerability markers rather than downstream consequences of illness, and may emerge early in the illness trajectory. Several immune markers showing causal effects on GMV, including IL-5 and leptin, are known mediators of immune activation and vascular instability, providing a biologically plausible link between peripheral inflammation, BBB dysfunction, and focal cortical thinning (Benest et al., Reference Benest, Kruse, Savant, Thomas, Laib, Loos, Fiedler and Augustin2013; Lisowska, Reference Lisowska2025; Wu & Zhang, Reference Wu and Zhang2023; Zou et al., Reference Zou, Zhong, Zhu, Zhao, Zhao, Cui, Gao and Li2019).
Interestingly, the MR analyses showed only partial overlap with the associations observed in our clinical cohort. This discrepancy may partly reflect differences in ancestry between the GWAS datasets and our clinical cohort. Moreover, it is not unexpected because MR estimates the effects of lifelong genetically determined immune variation, whereas observational analyses capture dynamic immune alterations influenced by disease status and environmental factors (Davies et al., Reference Davies, Holmes and Davey Smith2018). These approaches may provide complementary insights into neuroimmune mechanisms underlying MDD. Notably, IL-5 emerged as a convergent finding, showing a significant negative association with posterior cingulate cortex (PCC) volume. As a Th2-associated cytokine involved in adaptive immune regulation, IL-5 has been reported to be elevated in patients with MDD (Annam et al., Reference Annam, Galfalvy, Keilp, Simpson, Huang, Nandakumar, Byrnes, Nitahara, Hall, Stanley, Mann and Sublette2024), while structural abnormalities of the PCC are among the most consistently reported neuroanatomical features of the disorder. Given the central role of the PCC within the DMN and its involvement in emotional and self-referential processing, this finding may suggest a link between adaptive immune dysregulation and depression-related brain alterations (Bradley et al., Reference Bradley, Stern, Alonso, Xie, Kim-Schulze and Gabbay2019). One potential mechanism is that Th2-related immune dysregulation may increase blood–brain barrier permeability via endothelial tight junction impairment, facilitating peripheral immune signaling into the brain and thereby contributing to neuroinflammatory processes and structural vulnerability within limbic networks (Joo, Jeong, & Park, Reference Joo, Jeong and Park2023).
Peripheral immune and inflammatory alterations in MDD: state markers and biological heterogeneity
We observed significantly higher levels of IFN-γ, IL-4, IL-5, and TNF-β in patients with MDD compared with HCs, consistent with previous reports (Beurel, Toups, & Nemeroff, Reference Beurel, Toups and Nemeroff2020). In addition, elevated levels of chemokines such as MIP-1β and MIP-3α suggest a persistent dysregulation of chemokine-mediated affective regulation in MDD. This finding is consistent with experimental evidence showing that chemokine signaling within central circuits modulates fear- and anxiety-related behaviors (Choi et al., Reference Choi, Jeong, Kim, Lee, Ariyasiri, Kim, Jung, Hwang, Choi, Park, Huh, Carl, Rosenfeld, Raskin, Ma, Gecz, Kim, Kim, Shin and Kim2018). Notably, reductions in chemokine activity have been reported to accompany symptomatic improvement in MDD, indicating that the heightened chemokine profile observed in the present study may reflect an active pathological state rather than a trait marker (Bauer et al., Reference Bauer, Milenkovic, Hilbert, Sarubin, Weigl, Bahr, Wetter, Heckel, Wetzel, Rupprecht and Nothdurfter2020). Given the highly inconsistent evidence regarding MIF in MDD (Swoboda et al., Reference Swoboda, Deloch, von Zimmermann, Richter-Schmidinger, Lenz, Kornhuber and Mühle2022), the lower MIF levels observed in our cohort should be interpreted cautiously, and longitudinal studies are required to determine whether MIF reflects disease state or progression.
Stress-related processes in MDD have been linked to neurovascular dysfunction and increased BBB permeability, which is increasingly recognized as a core pathological feature of the disorder (Matsuno et al., Reference Matsuno, Tsuchimine, O’Hashi, Sakai, Hattori, Hidese, Nakajima, Chiba, Yoshimura, Fukuzato, Kando, Tatsumi, Ogawa, Ichinohe, Kunugi and Sohya2022). VEGF, a major regulator of angiogenesis and vascular permeability, has been implicated as a mediator of BBB disruption in depression (Lange et al., Reference Lange, Storkebaum, de Almodóvar, Dewerchin and Carmeliet2016). Angiopoietin-1 displays tightly regulated spatiotemporal expression patterns during brain development, supporting a genetically driven link between immune signaling, neurovascular regulation, and vulnerability to MDD (Wei, Jessa, Kleinman, & Phoenix, Reference Wei, Jessa, Kleinman and Phoenix2021).
In contrast, reduced TIMP-1 and IL-1R4 levels were observed in MDD. Given the critical role of TIMPs in the development of the central nervous system and neural plasticity, and evidence from related neuropsychiatric conditions showing reciprocal regulation of MMPs and TIMPs during clinical improvement, the reduced TIMP-1 levels observed in the present study may reflect illness-related downregulation of extracellular matrix remodeling pathways (Cirillo et al., Reference Cirillo, Pepe, Siciliano, Ippolito, Ricciardi, de Stefano, Buonanno, Atripaldi, Abbadessa, Perfetto, Sharbafshaaer, Sepe, Bonavita, Iavarone, Todisco, Papa, Tedeschi, Esposito and Trojsi2023). Additionally, IL-1R4 (ST2), a key component of the IL-33/ST2 signaling pathway, functions as an important regulator of immune homeostasis by coordinating innate and adaptive immune responses (Liew, Girard, & Turnquist, Reference Liew, Girard and Turnquist2016). It plays a critical role in limiting excessive inflammation and promoting tissue repair and immune resolution through IL-33–mediated signaling (Griesenauer & Paczesny, Reference Griesenauer and Paczesny2017). A previous study has suggested that dysregulation of inflammatory signaling pathways, including the IL-33/ST2 axis, is associated with depressive pathology, potentially contributing to sustained neuroinflammation and impaired neuroplasticity in MDD (Wu et al., Reference Wu, Zuo, Liu, Yang, Li and Li2026).
Structural brain vulnerability and immune-associated MDD subtypes
Notably, the structural connectivity abnormalities identified by IDSCN analysis in MDD were predominantly localized to the DMN and limbic network, with subcortical nuclei centered on the caudate serving as major affected regions. This pattern closely parallels the MR findings, which highlighted key cortico-striatal and limbic regions as being causally associated with MDD risk, indicating impairment of neural networks underlying cognitive and emotional regulation (Bliss-Moreau et al., Reference Bliss-Moreau, Santistevan, Bennett, Moadab and Amaral2021; Lemke et al., Reference Lemke, Romankiewicz, Förster, Meinert, Waltemate, Fingas, Grotegerd, Redlich, Dohm, Leehr, Thiel, Enneking, Brosch, Meller, Ringwald, Schmitt, Stein, Steinsträter, Bauer and Dannlowski2022; Zhang et al., Reference Zhang, Peng, Sweeney, Jia and Gong2018). Nevertheless, some discrepancies in the specific regions implicated by the two approaches were observed, which may partly reflect population-specific genetic architectures and racial differences between European and Asian cohorts (Fu et al., Reference Fu, Zhang, Wang, Wang, Zhang, Zhu, Qiu, Geng, Cui, Yu, Liao, Zhang, Gao, Xu, Han, Yao, Qin, Liu, Liang and Yu2024). Moreover, our findings suggest that MDD may be meaningfully stratified into two neurobiologically distinct subtypes. The hypoconnectivity subtype demonstrated significantly greater severity of anxiety symptoms and anhedonia relative to the hyperconnectivity subgroup. Functional MRI evidence indicates that individuals with anhedonia-predominant MDD exhibit hypoconnectivity within both the subcortical network and the dorsal attention network. This pattern of reduced functional connectivity may underlie core clinical features of MDD, including hyperarousal, impaired social motivation, and diminished hedonic capacity (Wang et al., Reference Wang, Tang, Zhang, Bu, Lu, Li, Gao, Hu, Kuang, Jia, Sweeney, Gong and Huang2021). Furthermore, prior neuroimaging research in MDD patients with comorbid elevated anxiety has reported attenuated functional connectivity in the DMN, implicating DMN dysregulation as a potential neural substrate of emotion dyscontrol, thereby contributing to symptom exacerbation and clinical instability (X. Wang et al., Reference Wang, Wei, Shao, Xue, Chen, Yao and Lu2025).
Furthermore, we demonstrated that the hyperconnectivity subtype was characterized by elevated levels of core inflammatory, chemokine-related, adaptive immune, and tissue repair markers, accompanied by reduced levels of fibrosis-associated factors. This immune signature is consistent with experimental evidence showing that increased IL-6, a central mediator of inflammatory signaling, enhances glutamatergic synaptic transmission and induces widespread brain hyperconnectivity that persists into adulthood in offspring (Mirabella et al., Reference Mirabella, Desiato, Mancinelli, Fossati, Rasile, Morini, Markicevic, Grimm, Amegandjin, Termanini, Peano, Kunderfranco, di Cristo, Zerbi, Menna, Lodato, Matteoli and Pozzi2021). Notably, a similar pattern of immune dysregulation has also been reported in hyperconnectivity-related subtypes of autism spectrum disorder, suggesting that aberrant peripheral immune activity may contribute to convergent alterations in cortical network organization across neuropsychiatric conditions (Pagani et al., Reference Pagani, Zerbi, Gini, Alvino, Banerjee, Barberis, Basson, Bozzi, Galbusera, Ellegood, Fagiolini, Lerch, Matteoli, Montani, Pozzi, Provenzano, Scattoni, Wenderoth, Xu and Gozzi2026).
Hyperconnectivity subtype exhibited a better response to antidepressant treatment
Furthermore, survival analysis revealed that the hyperconnectivity subtype with a higher inflammatory-immune state exhibited a faster response to antidepressant treatment. On one hand, consistent with previous studies (Korgaonkar, Goldstein-Piekarski, Fornito, & Williams, Reference Korgaonkar, Goldstein-Piekarski, Fornito and Williams2020; X. Wang et al., Reference Wang, Wei, Shao, Xue, Chen, Yao and Lu2025), the subtype characterized by enhanced connectivity between the DMN and limbic network, as well as within the limbic network itself, showed the most rapid antidepressant improvement. Previous evidence suggests that stronger intrinsic functional connectivity and network organization are associated with better antidepressant outcomes and response likelihood, particularly in circuits involved in emotional processing and regulation (Liston et al., Reference Liston, Chen, Zebley, Drysdale, Gordon, Leuchter, Voss, Casey, Etkin and Dubin2014). Higher baseline connectivity may reflect a neural architecture more capable of supporting treatment-related plasticity and clinical improvement. On the other hand, a high inflammatory/immune state is also a factor worth considering. Inflammatory biomarkers may indicate a distinct biological subgroup that responds to treatment, while also informing clinical stratification (Ha et al., Reference Ha, Kang, Lee, Kang, Kim and Lee2026). Elevated inflammatory states may reflect both increased vulnerability and a modifiable target, as standard antidepressants and adjunctive agents exert immunomodulatory effects (Więdłocha et al., Reference Więdłocha, Marcinowicz, Krupa, Janoska-Jaździk, Janus, Dębowska, Mosiołek, Waszkiewicz and Szulc2018). Anti-inflammatory augmentation strategies have shown preliminary benefit, particularly in inflamed subtypes (Kappelmann et al., Reference Kappelmann, Lewis, Dantzer, Jones and Khandaker2018), suggesting that immune load may function as a risk-enrichment marker with potential implications for closer monitoring and earlier treatment adjustment.
Beyond the biological implications, our findings also have potential relevance for emerging AI-driven brain research. Recent perspectives have proposed that global neuroscience would benefit from a standardized, large-scale, and interoperable ‘BrainNet’ framework, analogous to the transformative role of ImageNet in artificial intelligence, enabling multimodal data integration and reproducible model development (Zuo, Reference Zuo2026). In this context, genetically informed causal evidence linking immune dysregulation, regional brain structure, and MDD may provide biologically grounded features for future AI4Brain models (Richens, Lee, & Johri, Reference Richens, Lee and Johri2020). As harmonized neuroimaging, genomic, and immune datasets continue to expand, integrating causal inference with multimodal machine learning may facilitate more robust identification of biologically meaningful MDD subtypes and improve individualized risk prediction and treatment stratification (Koutsouleris et al., Reference Koutsouleris, Dwyer, Degenhardt, Maj, Urquijo-Castro, Sanfelici, Popovic, Oeztuerk, Haas, Weiske, Ruef, Kambeitz-Ilankovic, Antonucci, Neufang, Schmidt-Kraepelin, Ruhrmann, Penzel, Kambeitz and Haidl2021).
Several limitations should be noted. First, inflammatory markers were not repeatedly assessed during follow-up, limiting characterization of longitudinal immune dynamics under treatment. Longitudinal immune profiling in future studies will be necessary to determine whether immune dynamics track clinical improvement. Second, only peripheral immune measures were available, which may not fully capture central neuroimmune activity, given potential modulation by BBB permeability and neuroendocrine pathways. Third, unmeasured confounders, including lifestyle factors, metabolic status, or chronic medical conditions, may concurrently affect cortical structure and immune function. In addition, the neuroimaging and genetic priors underlying our analyses were largely derived from European-ancestry GWAS datasets, which may limit generalizability to non-European populations, including our Chinese cohort, and caution is needed when extrapolating these brain–immune associations across ancestries. Finally, due to the limited number of remitters, we were unable to examine differential treatment responses by medication class or dosage, and thus could not determine whether the identified immune and neurostructural markers differentially predict responses to treatments with distinct mechanisms of action, which warrants further investigation.
Conclusions
In conclusion, this study integrates MR, neuroimaging, immune profiling, and longitudinal follow-up to characterize neuroimmune heterogeneity in MDD. Beyond identifying brain regions genetically associated with MDD risk, we demonstrate bidirectional associations between peripheral immune markers and brain structural networks, supporting a genetically informed bidirectional framework linking immune and brain systems in depression vulnerability. Within MDD, abnormalities primarily involving connectivity between the DMN and limbic network, as well as connectivity within the limbic network, delineated biologically distinct subtypes with divergent inflammatory profiles and treatment trajectories. Notably, the subtype characterized by enhanced connectivity and heightened inflammatory activity exhibited a more rapid antidepressant response. Together, these findings suggest that immune dysregulation and large-scale brain network organization represent complementary dimensions of MDD heterogeneity and may inform biologically grounded patient stratification and personalized treatment strategies.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726105789.
Data availability statement
The GWAS data of MDD are available from the PGC database: https://pgc.unc.edu/for-researchers/download-results. The GWAS data of GMV are available from the UK Biobank Brain Imaging database: https://www.fmrib.ox.ac.uk/ukbiobank. The MATLAB code used in this study for IDSCN construction and clustering analyses is publicly available on GitHub at: https://github.com/willeydu/IDSCNs_code.
Acknowledgments
We gratefully acknowledge the researchers and participants of the Psychiatric Genomics Consortium (PGC), as well as the GWAS resources. Additionally, we particularly thank Figdraw for the assistance in creating Figure 1. The computations in this paper were supported by the High Performance Computing platform at West China Biomedical Big Data Center, West China Hospital, Sichuan University. The REST-meta-MDD consortium was initiated and established by Dr Chaogan Yan, who came from the Institute of Psychology, Chinese Academy of Sciences. This consortium aimed to address limited statistical power and analytic heterogeneity of major depressive disorder. For this purpose, 25 research groups from 17 hospitals in China formed the REST-meta-MDD consortium and agreed to share final resting-state fMRI indices and the structural data of patients with depression and matched normal controls. Detailed information can be found in http://rfmri.org/REST-meta-MDD. Researchers in the REST-meta-MDD consortium include Chao-Gan Yan, Xiao Chen, Le Li, Francisco Xavier Castellanos, Tong-Jian Bai, Qi-Jing Bo, Jun Cao, Guan-Mao Chen, Ning-Xuan Chen, Wei Chen, Chang Cheng, Yu-Qi Cheng, Xi-Long Cui, Jia Duan, Yi-Ru Fang, Qi-Yong Gong, Wen-Bin Guo, Zheng-Hua Hou, Lan Hu, Li Kuang, Feng Li, Kai-Ming Li, Tao Li, Yan-Song Liu, Zhe-Ning Liu, Yi-Cheng Long, Qing-Hua Luo, Hua-Qing Meng, Dai-Hui Peng, Hai-Tang Qiu, Jiang Qiu, Yue-Di Shen, Yu-Shu Shi, Chuan-Yue Wang, Fei Wang, Kai Wang, Li Wang, Xiang Wang, Ying Wang, Xiao-Ping Wu, Xin-Ran Wu, Chun-Ming Xie, Guang-Rong Xie, Hai-Yan Xie, Peng Xie, Xiu-Feng Xu, Hong Yang, Jian Yang, Jia-Shu Yao, Shu-Qiao Yao, Ying-Ying Yin, Yong-Gui Yuan, Ai-Xia Zhang, Hong Zhang, Ke-Rang Zhang, Lei Zhang, Zhi-Jun Zhang, Ru-Bai Zhou, Yi-Ting Zhou, Jun-Juan Zhu, Chao-Jie Zou, Tian-Mei Si, Xi-Nian Zuo, Jing-Ping Zhao, and Yu-Feng Zang.
Author contribution
W.D.: Writing – original draft, Investigation, Methodology, Visualization, Formal analysis. M.Y.: Writing – original draft, Investigation, Methodology, Visualization, Formal analysis. Z.G.: Methodology, Visualization, Formal analysis, Funding acquisition. H.L.: Methodology, Formal analysis, Software. W.X.: Data curation, Methodology, Formal analysis. G.L.: Data curation, Methodology, Formal analysis. E.L.: Data curation, Formal analysis, Investigation. D.J.H.: Writing – review and editing, Methodology, Validation. Y.K.: Data curation, Resources, Investigation. D.S.: Data curation, Resources, Investigation. J.C.: Data curation, Resources, Investigation. H.Z.: Data curation, Resources, Supervision. S.L.: Writing – review and editing, Project administration, Resources, Supervision, Funding acquisition. W.Z.: Writing – review and editing, Data curation, Project administration, Resources, Supervision, Funding acquisition.
Funding statement
This study was supported by the National Natural Science Foundation of China (Project Nos. 82401796, 82441007, and 82120108014), Department of Science and Technology of Sichuan Province (Project Nos. 2025ZNSFSC1568 and 2025ZNSFSC1771), 1·3·5 project for disciplines of excellence, West China Hospital, Sichuan University (Project Nos. ZYJC21004 and ZYGD23003), and 2025 Guiding Special Program of the Engineering Research Centre of Medical Information Technology, Ministry of Education (Project No. SCU2025D010).
Competing interests
The authors declare none.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the Ethics Committee of West China Hospital, Sichuan University (No. #2020:1054) on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
