Introduction
Major depressive disorder is one of the leading causes of disability globally, affecting approximately 350 million people annually (Murray et al., Reference Murray, Vos, Lozano, Naghavi, Flaxman, Michaud, Ezzati, Shibuya, Salomon, Abdalla, Aboyans, Abraham, Ackerman, Aggarwal, Ahn, Ali, Alvarado, Anderson, Anderson and Memish2012; Schmaal et al., Reference Schmaal, Hibar, Sämann, Hall, Baune, Jahanshad, Cheung, van Erp, Bos, Ikram, Vernooij, Niessen and Tiemeier2017a). Despite extensive research, understanding the underlying mechanisms of MDD remains challenging due to significant heterogeneity among individuals with the same diagnosis (Beijers, Wardenaar, van Loo, & Schoevers, Reference Beijers, Wardenaar, van Loo and Schoevers2019b). Two individuals with MDD may present with vastly different symptom profiles, course trajectories, and treatment responses (Hasler, Reference Hasler2010; Krishnan & Nestler, Reference Krishnan and Nestler2008). This considerable variability supports the notion that MDD encompasses multiple subtypes, though their neurobiological foundations are still not well understood, impeding the development of precise clinical guidelines. Clinically, MDD patients are often categorized into distinct subtypes based on symptom presentation (Lynch, Gunning, & Liston, Reference Lynch, Gunning and Liston2020). However, such symptom-based classifications have not led to effective, evidence-based treatment options. MDD patients with similar symptoms may still harbor different biological mechanisms, limiting the utility of these categories for differentiating etiologically distinct groups.
An alternative approach is to identify neurophysiological subtypes through clustering individuals based on objective neuroimaging signatures (Clementz et al., Reference Clementz, Sweeney, Hamm, Ivleva, Ethridge, Pearlson, Keshavan and Tamminga2016; Drysdale et al., Reference Drysdale, Grosenick, Downar, Dunlop, Mansouri, Meng, Fetcho, Zebley, Oathes, Etkin, Schatzberg, Sudheimer, Keller, Mayberg, Gunning, Alexopoulos, Fox, Pascual-Leone, Voss and Liston2017). This method has begun to shed light on how diverse biological mechanisms may contribute to the overlapping and heterogeneous clinical manifestations of psychiatric disorders (Clementz et al., Reference Clementz, Sweeney, Hamm, Ivleva, Ethridge, Pearlson, Keshavan and Tamminga2016; Hill, Reilly, Keefe, Gold, & Sweeney, Reference Hill, Reilly, Keefe, Gold and Sweeney2013). Structural neuroimaging studies using magnetic resonance imaging (MRI) have consistently reported abnormalities in cortical and subcortical regions, including the orbitofrontal cortex, superior frontal cortex, and hippocampus (Li et al., Reference Li, Zhao, Chen, Long, Dai, Huang, Lui, Radua, Vieta, Kemp, Sweeney, Li and Gong2020; 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 G.2017b; Schmaal et al., Reference Schmaal, Veltman, van Erp, Sämann, Frodl, Jahanshad, Loehrer, Tiemeier, Hofman, Niessen, Vernooij, Ikram, Wittfeld, Grabe, Block, Hegenscheid, Völzke, Hoehn, Czisch, Lagopoulos and Hibar2016). However, findings across studies have been inconsistent, largely due to the predominant use of case–control designs that overlook the interindividual heterogeneity within MDD (Belleau, Treadway, & Pizzagalli, Reference Belleau, Treadway and Pizzagalli2019; Bora, Fornito, Pantelis, & Yücel, Reference Bora, Fornito, Pantelis and Yücel2012; Chen et al., Reference Chen, Peng, Sun, Kuang, Li, Jia and Gong2016; Treadway et al., Reference Treadway, Waskom, Dillon, Holmes, Park, Chakravarty, Dutra, Polli, Iosifescu, Fava, Gabrieli and Pizzagalli2015). Contradictory and even opposing patterns of structural alterations have been observed in key regions such as the insula and anterior cingulate cortex (Peng et al., Reference Peng, Shi, Li, Fralick, Shen, Qiu, Liu, Jiang, Shen and Fang2015; 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 G.2017b; Zorlu, Cropley, Zorlu, Delibas, & Pantelis, Reference Zorlu, Cropley, Zorlu, Delibas, Adibelli, Baskin, Esen, Bora and Pantelis2017). These heterogeneous alterations may reflect distinct subtypes of MDD (Feder et al., Reference Feder, Sundermann, Wersching, Teuber, Kugel, Teismann, Heindel, Berger and Pfleiderer2017; Wang et al., Reference Wang, Tang, Zhang, Bu, Lu, Li, Gao, Hu, Kuang, Jia, Sweeney, Gong and Huang2021). Recently, the normative model has been proposed as a method to capture individual variability in structural abnormalities within psychiatric disorders (Marquand, Rezek, Buitelaar, & Beckmann, Reference Marquand, Rezek, Buitelaar and Beckmann2016; Wolfers et al., Reference Wolfers, Doan, Kaufmann, Alnæs, Moberget, Agartz, Buitelaar, Ueland, Melle, Franke, Andreassen, Beckmann, Westlye and Marquand2018). Similar to growth charts in pediatric medicine, which map child development based on height or weight, the normative model maps clinically relevant biological measures – such as neuroimaging metrics – using healthy control data (Marquand et al., Reference Marquand, Rezek, Buitelaar and Beckmann2016; Wolfers et al., Reference Wolfers, Doan, Kaufmann, Alnæs, Moberget, Agartz, Buitelaar, Ueland, Melle, Franke, Andreassen, Beckmann, Westlye and Marquand2018). Deviations from these normative models in brain structure and function show promise in identifying subtypes of psychiatric disorders, including MDD (Li et al., Reference Li, Long, Sheng, Du, Qiu, Chen and Liao2024a; Liu et al., Reference Liu, Jia, He, Wen, Zhang and Han2024; Sun et al., Reference Sun, Sun, Lu, Dong, Zhang, Wang, Liu, Ma, Wang, Wei, Chen, Liu, Huang, Zheng, Wu, Chen, Cheng, Xu, Gong and Xia2023a). However, most studies to date have been limited by small sample sizes and single-site designs, leaving open the question of whether individualized deviations from normative models of gray matter volume can identify robust and reproducible subtypes in multi-site neuroimaging datasets.
MDD arises from complex interactions across biological systems, spanning genes, molecules, cells, networks, and behavior (Anderson, Collins, Kong, Fang, & Holmes, Reference Anderson, Collins, Kong, Fang and Holmes2020). Increasing evidence suggests a strong relationship between gray matter morphological abnormalities and the brain’s genetic architecture and neurotransmitter receptor profiles. For instance, recent studies have demonstrated that gene expression profiles across the brain are closely associated with macroscopic neuroimaging phenotypes in healthy individuals (Krienen, Yeo, Ge, Buckner, & Sherwood, Reference Krienen, Yeo, Ge, Buckner and Sherwood2016; Li & Seidlitz, Reference Li and Seidlitz2021; Vértes et al., Reference Vértes, Rittman, Whitaker, Romero-Garcia, Váša and Kitzbichler2016) and altered neuroimaging metrics in brain disorders, as evidenced by the Allen Human Brain Atlas (AHBA), which provides gene expression data from six postmortem adult brains (Han et al., Reference Han, Zheng, Li, Zhou, Jiang, Wang, Wei, Pang, Li, Zhang, Chen and Cheng2022b; Hawrylycz et al., Reference Hawrylycz, Lein, Guillozet-Bongaarts, Shen, Ng, Miller, van de Lagemaat, Smith, Ebbert, Riley, Abajian, Beckmann, Bernard, Bertagnolli, Boe, Cartagena, Chakravarty, Chapin, Chong and Jones2012; Morgan, Seidlitz, & Whitaker, Reference Morgan, Seidlitz and Whitaker2019; Romme, de Reus, Ophoff, Kahn, & van den Heuvel, Reference Romme, de Reus, Ophoff, Kahn and van den Heuvel2017). Transcriptional similarity between brain regions – referred to as the correlated gene expression (CGE) connectome – correlates with structural and functional brain networks (Arnatkeviciute, Fulcher, Oldham, Tiego, & Fornito, Reference Arnatkeviciute, Fulcher, Oldham, Tiego and Fornito2021; Fulcher & Fornito, Reference Fulcher and Fornito2016). Recent study by Li et al. has shown that cortical thickness differences in MDD subtypes are constrained by CGE, with these subtypes exhibiting distinct CGE-informed epicenters, providing new evidence for a connectome-based spreading process in mental disorders (Hansen et al., Reference Hansen, Shafiei, Vogel, Smart, Bearden, Hoogman, Franke, van Rooij, Buitelaar, McDonald, Sisodiya, Schmaal, Veltman, van den Heuvel, Stein, van Erp, Ching, Andreassen, Hajek, Opel and Misic2022c; Li et al., Reference Li, Long, Sheng, Du, Qiu, Chen and Liao2024a; Yang et al., Reference Yang, Jian, Qiu, Zhang, Cheng, Ji, Li, Wang, Li and Li2021b). Additionally, neurotransmitter receptor profiles contribute to cortical morphology across various brain disorders (Hansen et al., Reference Hansen, Shafiei, Vogel, Smart, Bearden, Hoogman, Franke, van Rooij, Buitelaar, McDonald, Sisodiya, Schmaal, Veltman, van den Heuvel, Stein, van Erp, Ching, Andreassen, Hajek, Opel and Misic2022c). Understanding the transcriptional and molecular basis of gray matter abnormalities in MDD subtypes may enhance our understanding of their pathophysiology and help guide treatment strategies (Chen et al., Reference Chen, Müller, Dukart, Hoffstaedter, Baker, Holmes, Vatansever, Nickl-Jockschat, Liu, Derntl, Kogler, Jardri, Gruber, Aleman, Sommer, Eickhoff and Patil2021).
In this study, we aimed to identify potential subtypes of MDD based on individualized deviations from a normative model of gray matter volume in a large, multi-site cohort consisting of 1,276 MDD patients and 1,104 healthy controls (HCs) across 24 research sites. We first defined individualized deviations by comparing regional gray matter volumes in each patient to normative estimates derived from HCs. MDD subtypes were then identified using a K-means clustering algorithm and characterized in terms of clinical variables and voxel-wise gray matter abnormalities relative to HCs. Next, we predicted regional gray matter abnormalities based on deviations in neighboring regions as defined by the CGE connectome and explored potential epicenters within each MDD subtype. Finally, to gain a deeper understanding of the molecular mechanisms underlying these structural abnormalities, we examined the associations between the identified gray matter alterations and the distribution of neurotransmitter receptors.
Materials and methods
Imaging dataset and preprocessing
The MDD imaging datasets used in this study were sourced from the REST-meta-MDD consortium (http://rfmri.org/REST-meta-MDD) (Chen 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 Wang2019). The dataset includes data from 24 research sites, comprising 1,276 patients diagnosed with MDD (mean age: 36.23 ± 21.38 years; females: 63.71%) and 1,104 healthy controls (HCs) (mean age: 36.15 ± 24.55 years; females: 58.06%). All data collection procedures were approved by the local ethics committees at each research site. MDD diagnoses were made by experienced psychiatrists based on DSM-IV criteria. Depression severity in patients was assessed using the Hamilton Depression Rating Scale (HAMD) (Hamilton, Reference Hamilton1960). Among the MDD patients, 538 were first-episode, 282 were recurrent, 447 were medication-naive, and 408 were on medication. Further details on demographic characteristics, clinical assessments, and imaging acquisition protocols have been previously described (Chen 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 Wang2019).
Structural MRI data were preprocessed following standardized protocols across sites to minimize analytic variation. The specific preprocessing steps for structural MRI data are detailed in prior publications (Chen 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 Wang2019). For this study, we utilized gray matter volume (GMV) maps derived from voxel-based morphometry (VBM) analysis, which were made available by the REST-meta-MDD consortium (http://rfmri.org/REST-meta-MDD) (Ashburner, Reference Ashburner2009; Craddock, James, Holtzheimer, Hu, & Mayberg, Reference Craddock, James, Holtzheimer, Hu and Mayberg2012).
Normative model of gray matter volume and individualized gray matter morphological abnormalities
To mitigate non-biological variance introduced by differences in scanners and acquisition protocols, we applied the Combat method to harmonize GMV data across different sites. Combat has been demonstrated to effectively remove site-related variance while preserving biological variability in neuroimaging studies (Jean-Philippe et al., Reference Jean-Philippe, Drew, Birkan, Takanori, Mark and Neuroimage2017.
We then constructed a normative model of GMV. In line with methods outlined in previous research (Marquand et al., Reference Marquand, Rezek, Buitelaar and Beckmann2016; Wolfers et al., Reference Wolfers, Doan, Kaufmann, Alnæs, Moberget, Agartz, Buitelaar, Ueland, Melle, Franke, Andreassen, Beckmann, Westlye and Marquand2018), we trained a Gaussian process regression model to estimate the normative range of regional GMV values based on age and sex in HCs, using the Shen 268 brain atlas (Shen, Tokoglu, Papademetris, & Constable, Reference Shen, Tokoglu, Papademetris and Constable2013). This model was then applied to the MDD patient group. Individualized GMV abnormalities were represented as Z-scores, with a positive Z-score indicating a GMV value higher than the HC average, and a negative Z-score indicating a lower GMV value.
Before applying the model to the patient group, we assessed its performance using the following validation strategies: (1) 10-fold cross-validation, repeated 100 times; (2) leave-one-site-out cross-validation, where one site was used as the test set, and the remaining sites served as the training set in turn. Model performance was evaluated by calculating the standardized mean squared error (MSE) between the true and predicted GMV values (Marquand et al., Reference Marquand, Rezek, Buitelaar and Beckmann2016). After confirming the model’s validity, Z-scores were computed for each patient relative to the HCs, resulting in a Z-score matrix (patients × brain regions, N × 268).
To further investigate the intersubject heterogeneity of extreme deviations, we generated a spatial overlap map. This map quantified the percentage of participants exhibiting extreme deviations (defined as Z-scores exceeding ±1.96, representing a two-tailed 95% confidence interval). This approach allowed us to delineate extreme positive or negative deviations from the normative model (Haas et al., Reference Haas, Ge, Agartz, Amminger, Andreassen, Bachman, Baeza, Choi, Colibazzi, Cropley, de la Fuente-Sandoval, Ebdrup, Fortea, Fusar-Poli, Glenthøj, Glenthøj, Haut, Hayes, Heekeren and Frangou2024; Liu et al., Reference Liu, Jia, He, Wen, Zhang and Han2024; Sun, Sun, Lu, et al., Reference Sun, Sun, Lu, Dong, Zhang, Wang, Liu, Ma, Wang, Wei, Chen, Liu, Huang, Zheng, Wu, Chen, Cheng, Xu, Gong and Xia2023a).
Identifying MDD subtypes using individualized gray matter morphological abnormalities
We next sought to identify potential MDD subtypes based on individualized gray matter morphological abnormalities. To assess the significance of clustering, we first applied the SigClust approach. This method tests the hypothesis that the data fit a single Gaussian distribution, indicating no subtypes (Dinga et al., Reference Dinga, Schmaal, Penninx, Van Tol, Veltman, Van Velzen, Mennes, Van, Nic and Marquand2019). Upon confirming the presence of subtypes, the K-means algorithm was conducted to identify MDD subtypes based on individualized gray matter morphological abnormalities, using the correlation distance (one minus the correlation coefficients between samples) as the distance metric. To avoid local minima, K-means was run 100 times with different centroid initializations (Allen et al., Reference Allen, Damaraju, Plis, Erhardt, Eichele and Calhoun2014). The optimal number of subtypes, ranging from 2 to 10, was determined using a cluster ensemble voting technique, which incorporates 26 indices to ensure an unbiased selection of the optimal subtype (Charrad, Ghazzali, Boiteau, & Niknafs, Reference Charrad, Ghazzali, Boiteau and Niknafs2015).
Cluster stability was evaluated using two validation methods: subsample validation and leave-one-site-out validation. In the first approach, to minimize the influence of any small subset of patients, 90% of the participants were randomly selected, and k-means clustering was performed on this subsample. The Adjusted Rand Index (ARI) was calculated to compare the clustering results from the subsample with those derived from the full sample, assessing the consistency across iterations. This process was repeated 100 times to ensure robustness. In the second method, K-means clustering was performed excluding one site at a time, and the ARI between the clustering results obtained from each leave-one-site-out approach and the full-site clustering was calculated.
Characterizing subtypes
The identified subtypes of MDD were characterized based on demographic and clinical variables including sex, age, illness duration, symptom severity, and the distribution of first-episode versus recurrent, and untreated versus treated patients. Statistical comparisons were made using appropriate tests, including t-tests for continuous variables and chi-square tests for categorical variables.
Voxel-wise gray matter morphological differences between each MDD subtype and HCs were assessed using two-sample t-tests in SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/), with sex, age, and site as covariates. Additionally, voxel-wise gray matter abnormalities for the entire patient cohort were compared with HCs. Statistical significance was set at p < 0.05 with Bonferroni correction (family wise error p FWE < 0.05).
Associations between neurotransmitter receptors/transporters and gray matter morphological abnormalities of the identified subtypes
We examined the relationships between neurotransmitter receptors/transporters and gray matter morphological abnormalities of the identified subtypes. The data on neurotransmitter receptors/transporters were sourced from the PET-derived atlas compiled by Hansen et al. (Hansen & Shafiei, Reference Hansen and Shafiei2022b). This atlas includes several receptors and transporters: serotonin (5HT1A [Savli et al., Reference Savli, Bauer, Mitterhauser, Ding, Hahn, Kroll, Neumeister, Haeusler, Ungersboeck, Henry, Isfahani, Rattay, Wadsak, Kasper and Lanzenberger2012], 5HT1B [Baldassarri et al., Reference Baldassarri, Hillmer, Anderson, Jatlow, Nabulsi, Labaree, Cosgrove, O’Malley, Eissenberg, Krishnan-Sarin and Esterlis2018; Gallezot et al., Reference Gallezot, Nabulsi, Neumeister, Planeta-Wilson, Williams, Singhal, Kim, Maguire, McCarthy, Frost, Huang, Ding and Carson2010; Matuskey et al., Reference Matuskey, Bhagwagar, Planeta, Pittman, Gallezot, Chen, Wanyiri, Najafzadeh, Ropchan, Geha, Huang, Potenza, Neumeister, Carson and Malison2014; Murrough et al., Reference Murrough, Czermak, Henry, Nabulsi, Gallezot, Gueorguieva, Planeta-Wilson, Krystal, Neumaier, Huang, Ding, Carson and Neumeister2011a; Murrough et al., Reference Murrough, Henry, Hu, Gallezot, Planeta-Wilson, Neumaier and Neumeister2011b; Saricicek et al., Reference Saricicek, Chen, Planeta, Ruf, Subramanyam, Maloney, Matuskey, Labaree, Deserno, Neumeister, Krystal, Gallezot, Huang, Carson and Bhagwagar2015; Savli et al., Reference Savli, Bauer, Mitterhauser, Ding, Hahn, Kroll, Neumeister, Haeusler, Ungersboeck, Henry, Isfahani, Rattay, Wadsak, Kasper and Lanzenberger2012], 5HT2A [Beliveau & Ganz, Reference Beliveau and Ganz2017], 5HT4 [Beliveau & Ganz, Reference Beliveau and Ganz2017], 5HT6 [Radhakrishnan et al., Reference Radhakrishnan, Matuskey, Nabulsi, Gaiser, Gallezot, Henry, Planeta, Lin, Ropchan, Huang, Carson and D’Souza2020; Radhakrishnan et al., Reference Radhakrishnan, Nabulsi, Gaiser, Gallezot, Henry, Planeta, Lin, Ropchan, Williams, Morris, D’Souza, Huang, Carson and Matuskey2018], 5HTT [Beliveau & Ganz, Reference Beliveau and Ganz2017]), norepinephrine (α4β2 [Baldassarri et al., Reference Baldassarri, Hillmer, Anderson, Jatlow, Nabulsi, Labaree, Cosgrove, O’Malley, Eissenberg, Krishnan-Sarin and Esterlis2018; Hillmer et al., Reference Hillmer, Esterlis, Gallezot, Bois, Zheng, Nabulsi, Lin, Papke, Huang, Sabri, Carson and Cosgrove2016], M1 [Naganawa et al., Reference Naganawa, Nabulsi, Henry, Matuskey, Lin, Slieker, Schwarz, Kant, Jesudason, Ruley, Navarro, Gao, Ropchan, Labaree, Carson and Huang2021], VAChT [Aghourian, Legault-Denis, Soucy, & Rosa-Neto, Reference Aghourian, Legault-Denis, Soucy and Rosa-Neto2017; Bedard et al., Reference Bedard, Aghourian, Legault-Denis, Postuma, Soucy, Gagnon, Pelletier and Montplaisir2019]), cannabinoid (CB1 [D’Souza et al., Reference D’Souza, Cortes-Briones, Ranganathan, Thurnauer, Creatura, Surti, Planeta, Neumeister, Pittman, Normandin, Kapinos, Ropchan, Huang, Carson and Skosnik2016; Neumeister et al., Reference Neumeister, Normandin, Murrough, Henry, Bailey, Luckenbaugh, Tuit, Zheng, Galatzer-Levy, Sinha, Carson, Potenza and Huang2012; Normandin et al., Reference Normandin, Zheng, Lin, Mason, Lin, Ropchan, Labaree, Henry, Williams, Carson, Neumeister and Huang2015; Ranganathan et al., Reference Ranganathan, Cortes-Briones, Radhakrishnan, Thurnauer, Planeta, Skosnik, Gao, Labaree, Neumeister, Pittman, Surti, Huang, Carson and D’Souza2016]), dopamine (D1 [Kaller et al., Reference Kaller, Rullmann, Patt, Becker, Luthardt, Girbardt, Meyer, Werner, Barthel, Bresch, Fritz, Hesse and Sabri2017], D2 [Sandiego et al., Reference Sandiego, Gallezot, Lim, Ropchan, Lin, Gao, Morris and Cosgrove2015; Slifstein et al., Reference Slifstein, van de Giessen, Van Snellenberg, Thompson, Narendran, Gil, Hackett, Girgis, Ojeil, Moore, D’Souza, Malison, Huang, Lim, Nabulsi, Carson, Lieberman and Abi-Dargham2015; Smith et al., Reference Smith, Crawford, Dang, Seaman, San Juan, Vijay, Katz, Matuskey, Cowan, Morris, Zald and Samanez-Larkin2019; Zakiniaeiz et al., Reference Zakiniaeiz, Hillmer, Matuskey, Nabulsi, Ropchan, Mazure and Picciotto2019], DAT [Dukart et al., Reference Dukart, Holiga, Chatham, Hawkins, Forsyth, McMillan, Myers, Lingford-Hughes, Nutt, Merlo-Pich, Risterucci, Boak, Umbricht, Schobel, Liu, Mehta, Zelaya, Williams, Brown and Sambataro2018]), GABA (GABAa [Nørgaard et al., Reference Nørgaard, Beliveau, Ganz, Svarer, Pinborg, Keller, Jensen, Greve and Knudsen2021]), histamine (H3 [Gallezot et al., Reference Gallezot, Planeta, Nabulsi, Palumbo, Li, Liu, Rowinski, Chidsey, Labaree, Ropchan, Lin, Sawant-Basak, McCarthy, Schmidt, Huang and Carson2017], glutamate (mGluR5 [DuBois et al., Reference DuBois, Rousset, Rowley, Porras-Betancourt, Reader, Labbe, Massarweh, Soucy, Rosa-Neto and Kobayashi2016; Smart et al., Reference Smart, Cox, Scala, Tippler, Jaworska, Boivin, Séguin, Benkelfat and Leyton2019], NMDA [Galovic et al., Reference Galovic, Erlandsson, Fryer, Hong, Manavaki, Sari, Chetcuti, Thomas, Fisher, Sephton, Canales, Russell, Sander, Årstad, Aigbirhio, Groves, Duncan, Thielemans, Hutton and Koepp2021; McGinnity et al., Reference McGinnity, Hammers, Riaño Barros, Luthra, Jones, Trigg, Micallef, Symms, Brooks, Koepp and Duncan2014], opioid (MOR [Kantonen et al., Reference Kantonen, Karjalainen, Isojärvi, Nuutila, Tuisku, Rinne, Hietala, Kaasinen, Kalliokoski, Scheinin, Hirvonen, Vehtari and Nummenmaa2020]), and norepinephrine (NET [Belfort-DeAguiar et al., Reference Belfort-DeAguiar, Gallezot, Hwang, Elshafie, Yeckel, Chan, Carson, Ding and Sherwin2018; Ding et al., Reference Ding, Singhal, Planeta-Wilson, Gallezot, Nabulsi, Labaree, Ropchan, Henry, Williams, Carson, Neumeister and Malison2010; Li et al., Reference Li, Potenza, Lee, Planeta, Gallezot, Labaree, Henry, Nabulsi, Sinha, Ding, Carson and Neumeister2014; Sanchez-Rangel et al., Reference Sanchez-Rangel, Gallezot, Yeckel, Lam, Belfort-DeAguiar, Chen, Carson, Sherwin and Hwang2020]). PET images from each study were averaged across participants, normalized to the MNI-ICBM 152 nonlinear 2009 template, and subsequently parcellated into 268 brain regions. The average receptor/transport densities were Z-scored (Hansen & Shafiei, Reference Hansen and Shafiei2022a, Reference Hansen and Shafiei2022b). To assess the regional gray matter morphological abnormalities, unthresholded t-statistics for each region were averaged. A multilinear regression model was constructed for each subtype, incorporating neurotransmitter receptor/transport profiles and regional gray matter abnormalities. The statistical significance of these models was determined through permutation testing (10,000 iterations), with Bonferroni correction applied to adjust for multiple comparisons.
Subsequently, a dominance analysis was conducted for each model to evaluate the relative contribution of each predictor (i.e., neurotransmitter receptor/transport profiles) to the overall model fit (Budescu & David, Reference Budescu and David1993). Dominance analysis estimates the importance of predictors by applying the multilinear model across all possible combinations of predictors. In this study, the total dominance value of each predictor was derived from the average increase in explained variance (R2) when it was added to the sub-models (Azen & Budescu, Reference Azen and Budescu2003; Budescu & David, Reference Budescu and David1993; Hansen & Shafiei, Reference Hansen and Shafiei2022a).
Disease epicenter of correlated gene expression connectome for each subtype
To explore the genetic underpinnings of gray matter morphological abnormalities in each subtype, we analyzed the associations between the CGE connectome and regional gray matter morphology. The CGE connectome was derived from regional gene expression data obtained from the AHBA (http://human.brain-map.org) (Hawrylycz et al., Reference Hawrylycz, Lein, Guillozet-Bongaarts, Shen, Ng, Miller, van de Lagemaat, Smith, Ebbert, Riley, Abajian, Beckmann, Bernard, Bertagnolli, Boe, Cartagena, Chakravarty, Chapin, Chong and Jones2012). The raw expression data were preprocessed following established protocols (Fang, Hou, Niu, Han, & Zhang, Reference Fang, Hou, Niu, Han and Zhang2024; Hawrylycz et al., Reference Hawrylycz, Lein, Guillozet-Bongaarts, Shen, Ng, Miller, van de Lagemaat, Smith, Ebbert, Riley, Abajian, Beckmann, Bernard, Bertagnolli, Boe, Cartagena, Chakravarty, Chapin, Chong and Jones2012) and transcriptome similarities were quantified using Pearson’s correlation coefficients between regional gene expressions, resulting in a symmetric CGE connectome (268 × 268). Only positive correlations in the CGE connectome were retained for further analysis.
The relationship between the CGE connectome and regional gray matter abnormalities was assessed using methods previously outlined in the literature (Shafiei et al., Reference Shafiei, Markello, Makowski, Talpalaru, Kirschner, Devenyi, Guma, Hagmann, Cashman, Lepage, Chakravarty, Dagher and Misic2020a). For each subtype, we calculated the normalized collective differences of transcriptomic neighbors for each region (
$ {D}_i^{\mathrm{CGE}} $
) as described in the following:
$$ {D}_i^{\mathrm{sc}}=\frac{1}{N_i}{\sum}_{j=1,j\ne i}^{N_i}{T}_i\times {\mathrm{D}}_{i,j} $$
where
$ {D}_i^{\mathrm{CGE}} $
represents the normalized collective abnormalities of the transcriptomic neighbors for region i,
$ {D}_i $
is the regional abnormalities (unthresholded t-statistic) of region i,
$ {N}_i $
is the number of neighboring regions with transcriptomic connections, and
$ {\mathrm{D}}_{i,j} $
is the strength of the transcriptomic connection between region i and j. For each subtype, the
$ {D}_i^{CGE} $
was predicted based on the regional abnormalities of neighboring regions. The Pearson’s correlation coefficient between true abnormalities and
$ {D}_i^{CGE} $
across all brain regions was then calculated. To assess whether gray matter abnormalities specific to the identified subtypes exhibited a stronger association with the CGE connectome compared to all patients, we conducted a similar analysis for the entire patient cohort. The Pearson’s correlation coefficients were compared using Steiger’s Z test (Feng et al., Reference Feng, Yuan, Geng, Gu, Zhou, Wu and Luo2018; Fang et al., Reference Fang, Han, Li, Ding, Wu and Zhang2021).
Next, we identified potential disease epicenters for each subtype. A brain region was considered a disease epicenter if, along with its connected neighbors, it showed significantly greater abnormalities than other regions (Shafiei et al., Reference Shafiei, Markello, Makowski, Talpalaru, Kirschner, Devenyi, Guma, Hagmann, Cashman, Lepage, Chakravarty, Dagher and Misic2020a). Brain regions were ranked based on the differences in unthresholded t-statistics and CGE-informed abnormalities (
$ {D}_i^{\mathrm{sc}} $
) in ascending order. The average ranking values were then used to determine the disease epicenter likelihood rankings, with statistical significance evaluated through permutation testing (10,000 iterations). A brain region was designated as a disease epicenter if its likelihood ranking was significantly higher than expected by chance (permutation p < 0.01).
Reproducibility analysis
To ensure the robustness of our findings and assess the impact of brain parcellation choices, we validated our main results using an alternative brain atlas, the Automated Anatomical Labeling (AAL) atlas, which operates at a different resolution (Tzourio-Mazoyer et al., Reference Tzourio-Mazoyer, Landeau, Papathanassiou, Crivello, Etard, Delcroix, Mazoyer and Joliot2002).
Results
Clinical demographics
The clinical and demographic information for the datasets utilized in this study is presented in Supplementary Table S1.
Heterogeneous individualized extreme deviations in patients with MDD
The results of the 10-fold cross-validation and leave-one-site-out cross-validation demonstrated the high generalizability of the normal models of GMVs in predicting GMVs in HCs. The mean MSE for the 10-fold and leave-one-site-out cross-validation was 0.73 (±0.08) and 0.82 (±0.09), respectively. The spatial distributions of the MSE are shown in Supplementary Figure S1.
After confirming the performance of the normative model, we applied the trained model derived from HCs to patients with MDD to identify individualized gray matter morphological abnormalities. MDD patients exhibited heterogeneous patterns of abnormal gray matter morphology. Specifically, 59.01% (n = 753) of MDD patients showed significant negative deviations from the normative model in at least one brain region (Figure 1). These negative deviations were primarily found in the superior frontal gyrus, calcarine, dorsolateral prefrontal cortex (DLPFC), superior parietal cortex, and middle occipital cortex (Figure 1). Conversely, 50.71% (n = 647) of MDD patients showed extreme positive deviations in at least one brain region (Figure 1), most notably in the right inferior frontal gyrus, right amygdala, hippocampus, superior parietal cortex, and middle occipital cortex (Figure 1). Notably, the percentage of patients exhibiting extreme deviations from the normative range in any single brain region was relatively low, with the most common extreme deviation shared by only 5.64% of patients (Figure 1). These findings highlight substantial heterogeneity in individualized deviations from the normative model among MDD patients, underscoring the need to identify more homogeneous MDD subtypes.
Spatial overlap maps of extreme deviations in MDD patients. The left panel illustrates the overlap of extreme negative deviations (Z-scores <−1.96), and the right panel shows the overlap of extreme positive deviations (Z-scores >1.96) across MDD patients.
Two subtypes manifesting opposite patterns of gray matter morphological abnormalities are identified
Significant cluster testing confirmed the existence of distinct MDD subtypes characterized by individualized gray matter abnormalities (Supplementary Figure S2). The cluster ensemble voting technique identified a two-cluster solution as optimal for the entire MDD sample (Figure 2a). Patients were subsequently classified into two subtypes via K-means clustering (Subtype 1, N = 615; Subtype 2, N = 661). Subsample validation and leave-one-site-out validation results confirmed the robustness of these clustering outcomes, with the mean adjusted Rand index (ARI) between validation results and the primary results being 0.97 (±0.01) and 0.99 (±0.01), respectively. The distributions of ARI values are shown in Figure 2b.
Subtyping results. (a) The optimal two-cluster solution is identified by the cluster ensemble voting technique (marked by a red asterisk). (b) Subsample validation and leave-one-site-out (LOSO) validation results. The adjusted Rand Index (ARI) values between clustering outcomes from validation subsamples and the main results are shown. (c) Voxel-wise gray matter abnormalities in the identified subtypes relative to healthy controls.
Voxel-wise analysis of gray matter morphology revealed that the two subtypes displayed opposing patterns of deviations from HC GMVs with spatial Pearson’s correlation of −0.36 (p < 0.01). Specifically, subtype 1 showed significantly increased GMVs in regions such as the inferior and superior frontal gyrus, anterior cingulate cortex, insula, superior temporal gyrus, middle occipital gyrus, caudate, and inferior parietal lobule, while GMVs were decreased in cerebellar regions (Supplementary Table S1 and Figure 2c). In contrast, subtype 2 exhibited decreased GMVs in several brain regions, including the midline frontocortical areas (frontal gyrus, cingulate cortex), insula, and striatum, with increased GMVs in cerebellar regions (Supplementary Table S1 and Figure 2c). Despite these differences, the two subtypes did not show significant variation in clinical and demographic variables, including sex, age, illness duration, symptom severity, and the proportion of first-episode versus recurrent, or untreated versus treated patients (all uncorrected p > 0.05).
Gray matter morphological abnormalities in subtypes exhibit distinct associations with neurotransmitter receptors/transporters
We further explored the associations between neurotransmitter receptors/transporters and gray matter morphological abnormalities within each subtype. This was accomplished by constructing multilinear models to account for the spatial distributions of receptors/transporters and the differential patterns of gray matter abnormalities. The goodness-of-fit (adjusted R 2) of these models was 0.88 (F-statistic (268,248) = 101.00, p_FWE < 0.01) for subtype 1 and 0.90 (F-statistic (268,248) = 128.00, p_FWE < 0.01) for subtype 2, indicating strong model performance (see Figure 3a). Dominance analysis revealed that specific neurotransmitter receptors were crucial for each subtype. Notably, H3 receptor was significant for subtype 1, while mGluR5 was prominent for subtype 2 (Figure 3b).
Association between gray matter morphological abnormalities of the identified subtypes and neurotransmitter receptor/transporter distribution. (a) Multilinear models assessing the association between neurotransmitter receptors/transporters and gray matter abnormalities of subtypes. The bar plot illustrates the goodness-of-fit (adjusted R 2) for each model, with Bonferroni correction (p FWE < 0.01). (b) Dominance analysis for determining the relative importance of predictors in each model. (c) Cumulative contributions of excitatory versus inhibitory receptors to gray matter abnormalities in the identified subtypes, with the grey line representing the identity line.
Furthermore, we categorized receptors into excitatory and inhibitory types and calculated their cumulative contributions to gray matter abnormalities. Our results indicated that the gray matter abnormalities in subtype 1 were primarily associated with excitatory receptors, while those in subtype 2 were predominantly linked to inhibitory receptors (Figure 3c).
Divergent CGE-informed disease epicenters in the identified subtypes
Both subtypes exhibited significant correlations between regional gray matter abnormalities and the regional differences in connected neighbors within the CGE connectome (Subtype 1: r = 0.71, p FWE < 0.01; Subtype 2: r = 0.73, p FWE < 0.01; see Figure 4a). These findings suggest that the gray matter abnormalities in both subtypes are influenced by the CGE connectome. Moreover, gray matter abnormalities in the identified subtypes showed significantly higher Pearson’s correlations with the CGE connectome than in the overall MDD patient group (Subtype 1 vs. all patients, Steiger’s Z = 5.98, p FWE < 0.01; Subtype 2 vs. all patients, Steiger’s Z = 6.68, p FWE < 0.01).
Association between gray matter abnormalities in the identified subtypes and transcriptomic similarity based on correlated gene expression (CGE) connectome. (a) Pearson’s correlation coefficient between the regional patterns of gray matter abnormalities and the CGE-informed differential pattern. (b) Putative disease epicenters for each subtype.
Distinct disease epicenters were identified for the two subtypes (Figure 4b). Subtype 1 showed epicenters in the cerebellar anterior lobe, hippocampus, thalamus, and posterior cingulate cortex (permutation p FWE < 0.01). Subtype 2, however, exhibited epicenters in regions such as the middle frontal gyrus, inferior parietal lobule extending to the superior temporal gyrus and cuneus (permutation p FWE < 0.01).
Reproducibility analysis
To assess the robustness of our findings across different brain parcellation schemes, we validated our main results using the AAL atlas. Significant cluster tests confirmed the presence of MDD subtypes (Supplementary Figure S3), and the cluster ensemble technique again identified the two-cluster solution as optimal for the MDD sample (Supplementary Figure S4A). The ARI values between the original and the AAL-based clustering outcomes were 0.77, indicating high consistency. The gray matter abnormalities of the identified subtypes remained largely unchanged (Supplementary Figure S4B). The spatial Spearman’s correlations between the abnormal patterns from the original and AAL-based analyses were 0.99 (p FWE < 0.01) for subtype 1 and 0.99 (p FWE < 0.01) for subtype 2, confirming the reproducibility of the identified subtypes.
Discussion
This study identified two robust and reproducible MDD subtypes based on individualized gray matter morphological abnormalities, derived from a large and multi-site MDD cohort. The identified subtypes show opposite patterns of gray matter morphological abnormalities compared to HCs. Specifically, we observed that subtype 1 is predominantly associated with excitatory receptor abnormalities, while subtype 2 is linked to inhibitory receptor dysfunctions. Moreover, the gray matter morphological abnormalities of the identified subtypes were constrained by CGE connectome and were related to distinct CGE-informed epicenter profiles. These findings contribute to a deeper understanding of the structural brain heterogeneity in MDD and suggest molecular mechanisms that may differentiate these subtypes.
Patients with MDD exhibit substantial individual variability in structural and functional neuroimaging features, including gray matter volume and functional connectome, to the extent that extreme person-specific deviations from normal expectation for regional GMV never exceed 6.21% of cases with MDD (Han et al., Reference Han, Cui, Zheng, Li, Zhou, Fang, Sheng, Wen, Liu, Wei, Chen, Chen, Cheng and Zhang2023a; Li et al., Reference Li, Long, Sheng, Du, Qiu, Chen and Liao2024b; Segal et al., Reference Segal, Parkes, Aquino, Kia, Wolfers, Franke, Hoogman, Beckmann, Westlye, Andreassen, Zalesky, Harrison, Davey, Soriano-Mas, Cardoner, Tiego, Yücel, Braganza, Suo and Fornito2023; Sun et al., Reference Sun, Sun, Lu, Dong, Zhang, Wang, Liu, Ma, Wang, Wei, Chen, Liu, Huang, Zheng, Wu, Chen, Cheng, Xu, Gong and Grp2023b). This heterogeneity is in line with the clinical heterogeneity, posing significant challenges for understanding the disorder’s etiology. Prior studies have shown inconsistent results regarding structural brain abnormalities, with some contradictory findings (Bora et al., Reference Bora, Fornito, Pantelis and Yücel2012; Chen et al., Reference Chen, Peng, Sun, Kuang, Li, Jia and Gong2016; Kempton et al., Reference Kempton, Salvador, Munafò, Geddes, Simmons, Frangou and Williams2011; Schmaal et al., Reference Schmaal, Veltman, van Erp, Sämann, Frodl, Jahanshad, Loehrer, Tiemeier, Hofman, Niessen, Vernooij, Ikram, Wittfeld, Grabe, Block, Hegenscheid, Völzke, Hoehn, Czisch, Lagopoulos and Hibar2016). The identification of homogeneous MDD subtypes is critical. While previous research has used data-driven approaches to classify MDD subtypes based on neuroimaging features, such studies typically relied on machine learning models applied to single, smaller cohorts (Beijers, Wardenaar, van Loo, & Schoevers, Reference Beijers, Wardenaar, van Loo and Schoevers2019a; Chen, Dai, & Lin, Reference Chen, Dai and Lin2023; Han et al., Reference Han, Xu, Guo, Fang, Wei, Liu, Cheng, Zhang and Cheng2022a; Xu et al., Reference Xu, Guo, Zheng, Wei, Wen, Fang, Zhang, Cheng and Han2023; Yang et al., Reference Yang, Kumar, Nickerson, Du, Wang, Chen, Li, Pizzagalli and Ma2021a; Yang et al., Reference Yang, Chen, Liu, Fan, Ouyang, Tao and Yang2025). In contrast, our approach demonstrates robust subtype identification, validated by subsample and leave-one-site-out validation, which supports the reproducibility of the findings. Furthermore, the two subtypes show opposite patterns of gray matter abnormalities relative to HCs. Specifically, subtype 1 exhibits increased GMVs in cortical regions and decreased GMVs in cerebellar areas, while subtype 2 shows the reverse pattern. A recent study identified similar subtypes based on cortical thickness patterns (Li et al., Reference Li, Long, Sheng, Du, Qiu, Chen and Liao2024a), and our results extend these findings to include subcortical and cerebellar regions, areas frequently implicated in depression pathophysiology (Gong & He, Reference Gong and He2015; Kaiser, Andrews-Hanna, Wager, & Pizzagalli, Reference Kaiser, Andrews-Hanna, Wager and Pizzagalli2015; Otte et al., Reference Otte, Gold, Penninx, Pariante, Etkin, Fava, Mohr and Schatzberg2016). The cerebellum, in particular, is of interest due to its anatomical connections to limbic regions (Turner et al., Reference Turner, Paradiso, Marvel, Pierson, Boles Ponto, Hichwa and Robinson2007) and its critical role in emotion and cognitive processing (Moulton et al., Reference Moulton, Elman, Pendse, Schmahmann, Becerra and Borsook2011; O’Reilly, Beckmann, Tomassini, Ramnani, & Johansen-Berg, Reference O’Reilly, Beckmann, Tomassini, Ramnani and Johansen-Berg2010), with interactions between the cerebellum and regions especially the hippocampus being central to depression (Steinke et al., Reference Steinke, Gaser, Langbein, Dietzek, Gussew, Reichenbach, Smesny, Sauer and Nenadić2017). However, both increased and atrophic cerebellum volume are reported previously (Chen et al., Reference Chen, Guo, Chen, Yang, Yan, Sun, Ma, Zhang, Qi, Fang, Jiang, Tao and Wang2024a; Chen et al., Reference Chen, Chen, Hu, Wang, Zhang, Lyu, Xu, Chen and Yu2025; Chen, Sui, Yang, Lv, & Wang, Reference Chen, Sui, Yang, Lv and Wang2024b). Despite prior reports of both enlarged and atrophic cerebellar volumes (Han et al., Reference Han, Zheng, Li, Liu, Wang, Jiang, Wen, Zhou, Wei, Pang, Li, Zhang, Chen and Cheng2023b), we found no significant differences in clinical variables such as illness duration or symptom severity between the subtypes, suggesting that the brain differences identified are not captured by conventional clinical measures and may have novel clinical implications.
To explore the biological mechanisms underlying the observed gray matter abnormalities, we examined their associations with neurotransmitter receptor profiles. Dysregulation of neurotransmitters is a key factor in the pathophysiology of MDD, and modern antidepressant treatments target specific neurotransmitter systems (Hansen & Shafiei, Reference Hansen and Shafiei2022a). Structural and functional abnormalities in psychiatric disorders have been linked to neurotransmitter receptor profiles (Hansen & Shafiei, Reference Hansen and Shafiei2022a, Reference Hansen and Shafiei2022b), and our findings corroborate this relationship. Specifically, subtype 1 is predominantly associated with the histamine H3 receptor, while subtype 2 is linked to the mGluR5 receptor. Both receptors have been implicated in MDD pathology; for example, histamine regulates various functions, including learning, memory, and appetite, and H3 receptor antagonists have shown pro-cognitive and antidepressant effects (Iida et al., Reference Iida, Yoshikawa, Kárpáti, Matsuzawa, Kitano, Mogi, Harada, Naganuma, Nakamura and Yanai2017; Femenía et al., Reference Femenía, Magara, DuPont and Lindskog2015). These receptor-specific associations may help explain why traditional antidepressant treatments often have limited efficacy (Möller, Reference Möller2008), highlighting the potential for more tailored interventions. Our findings provide new insights into the molecular bases underlying structural abnormalities in MDD subtypes, with implications for precision medicine.
Additionally, we find that gray matter abnormalities in the identified subtypes align with the CGE connectome, further supporting a network-based spread of structural abnormalities in MDD. Neuroimaging studies have demonstrated that these abnormalities often follow the underlying brain connectome rather than being randomly distributed, with structural changes originating from disease epicenters and propagating to other brain regions (Shafiei et al., Reference Shafiei, Markello, Makowski, Talpalaru, Kirschner, Devenyi, Guma, Hagmann, Cashman, Lepage, Chakravarty, Dagher and Mišić2020b; Wannan et al., Reference Wannan, Cropley, Chakravarty, Bousman, Ganella, Bruggemann, Weickert, Weickert, Everall, McGorry, Velakoulis, Wood, Bartholomeusz, Pantelis and Zalesky2019; Zhou, Gennatas, Kramer, Miller, & Seeley, Reference Zhou, Gennatas, Kramer, Miller and Seeley2012). This network-based hypothesis has been supported in MDD and other psychiatric disorders (Han, Cui, et al., Reference Han, Cui, Zheng, Li, Zhou, Fang, Sheng, Wen, Liu, Wei, Chen, Chen, Cheng and Zhang2023a; Li et al., Reference Li, Long, Sheng, Du, Qiu, Chen and Liao2024b). Moreover, gray matter morphology is highly heritable and constrained by genetic factors (Grasby et al., Reference Grasby, Jahanshad, Painter, Colodro-Conde, Bralten, Hibar, Lind, Pizzagalli, Ching, McMahon, Shatokhina, Zsembik, Thomopoulos, Zhu, Strike, Agartz, Alhusaini, Almeida, Alnæs and Medland2020; Panizzon et al., Reference Panizzon, Fennema-Notestine, Eyler, Jernigan, Prom-Wormley, Neale, Jacobson, Lyons, Grant, Franz, Xian, Tsuang, Fischl, Seidman, Dale and Kremen2009). Our results confirm that gray matter abnormalities are not only influenced by genetic architecture but also constrained by transcriptomic similarity, providing further evidence for the network-based spread of these abnormalities. The distinct epicenter patterns observed in our subtypes suggest divergent pathological trajectories, which could be further validated in longitudinal studies.
This study has several limitations. First, the cross-sectional design raises questions about the stability and generalizability of the identified subtypes. Future research should validate these findings with independent, longitudinal datasets to assess their robustness. Second, the limited clinical data available in this study constrained our ability to explore differences in other clinical variables, such as medication dosage or comorbidities. Future studies incorporating more comprehensive clinical data would provide deeper insights into subtype characteristics and their potential implications for treatment. Third, we did consider other neuroimaging metrics, such as intrinsic brain connectivity, which also provides a potentially mechanistic framework for understanding mental disorders (Cheng et al., Reference Cheng, Cai, Liu, Yang, Pan, Zhang, Mo, Yu and Zhu2025; Mo et al., Reference Mo, Zhao, Li, Cai, Song, Wang, Yu and Zhu2024; Xu et al., Reference Xu, Zhang, Zhou, Guo, Mo, Ma, Zhu and Qian2024; Zhang, Xu, Ma, Qian, & Zhu, Reference Zhang, Xu, Ma, Qian and Zhu2024). Finally, the impact of comorbidities on these results was not examined, and future study should investigate whether comorbid conditions influence the identified subtypes or contribute to additional subtyping within MDD.
In conclusion, we identified two distinct subtypes of MDD based on individualized gray matter morphological abnormalities, which exhibit contrasting patterns relative to the healthy population. These subtypes are associated with different CGE-informed epicenters and show divergent neurotransmitter receptor profiles, pointing to distinct molecular mechanisms underlying their structural brain differences. These findings contribute to our understanding of the heterogeneous nature of MDD and offer potential directions for more personalized therapeutic approaches.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291725101499.
Acknowledgements
This research study was supported by the Natural Science Foundation of China (62106229, 62476252), Medical Science and Technology Research Project of Henan Province (S20240045), Natural Science Foundation of Henan Province (252300421113), and Medical Science and Technology Research Project of Henan Province (LHGJ20230217).
Competing interests
All authors declared no conflict of interest.
Open access
