Introduction
Bipolar disorder (BD), major depressive disorder (MDD), and schizophrenia (SZ) collectively represent the second leading cause of disability worldwide and are associated with increased mortality from both natural causes and suicide (Global Burden of Disease, 2021). These complex and heterogeneous mental disorders are characterized by combinations of symptoms involving perceptions, emotions, thoughts, and behaviors. Current diagnostic criteria, which rely on partially overlapping symptom profiles, may result in diagnostic inaccuracies and suboptimal treatment outcomes (Richards et al., Reference Richards, Cardno, Harold, Craddock, Di Florio, Jones and O’Donovan2022). Over the past three decades, research in the field of brain mapping has sought to elucidate the pathophysiological mechanisms of these disorders to advance diagnostic, preventive, and therapeutic strategies. Evidence from structural magnetic resonance imaging (sMRI) technologies has highlighted consistent regional neuroanatomical variations across these clinical entities (Chen et al., Reference Chen, Wang, Gong, Qi, Fu, Tang and Wang2022; Gray, Müller, Eickhoff, & Fox, Reference Gray, Müller, Eickhoff and Fox2020; Liloia et al., Reference Liloia, Brasso, Cauda, Mancuso, Nani, Manuello and Rocca2021). While the current corpus of sMRI findings has deepened our understanding of the neural phenotype of BD, MDD, and SZ, the anticipated clinical applicability of structural neuroimaging remains unrealized (Abi-Dargham et al., Reference Abi-Dargham, Moeller, Ali, DeLorenzo, Domschke, Horga and Krystal2023; Stein et al., Reference Stein, Shoptaw, Vigo, Lund, Cuijpers, Bantjes and Maj2022).
One reason for this translational gap may lie in the tendency of mental illnesses to exhibit gray matter variations within a defined set of brain regions. Independent studies have identified abnormalities in hub nodes of the human connectome across distinct diagnostic entities (Cauda et al., Reference Cauda, Nani, Manuello, Liloia, Tatu, Vercelli and Costa2019; Crossley et al., Reference Crossley, Mechelli, Scott, Carletti, Fox, McGuire and Bullmore2014; Hettwer et al., Reference Hettwer, Larivière, Park, van den Heuvel, Schmaal, Andreassen and Valk2022; Kaczkurkin et al., Reference Kaczkurkin, Park, Sotiras, Moore, Calkins, Cieslak and Satterthwaite2019; Romer et al., Reference Romer, Elliott, Knodt, Sison, Ireland, Houts and Hariri2021; Vanasse et al., Reference Vanasse, Fox, Fox, Cauda, Costa, Smith and Lancaster2021; Writing Committee for the Attention-Deficit/Hyperactivity Disorder et al., 2021), raising the question of whether gray matter variations represent a disorder-specific neurobiological phenotype or reflect a transdiagnostic marker of psychiatric disorders (Bourque et al., Reference Bourque, Poulain, Proulx, Moreau, Joober, Forgeot d’Arc and Jacquemont2024; de Lange et al., Reference de Lange, Scholtens, van den Berg, Boks, Bozzali, Cahn and van den Heuvel2019; Segal et al., Reference Segal, Tiego, Parkes, Holmes, Marquand and Fornito2024). For example, the meta-analysis by Goodkind et al. (Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata and Etkin2015), encompassing 15,892 participants across six psychiatric diagnoses (i.e. MDD, BD, SZ, addiction, obsessive-compulsive disorder, and anxiety), identified a common cluster of gray matter reductions in the dorsal anterior cingulate cortex (dACC), anterior insula (AI), ventromedial and dorsomedial prefrontal cortex, thalamus, amygdala, hippocampus, superior temporal gyrus, and parietal operculum. Wise et al. (Reference Wise, Radua, Via, Cardoner, Abe, Adams and Arnone2017) further confirmed a shared neuroanatomical reduction in the dACC, AI, and the dorsomedial prefrontal cortex in individuals with MDD and BD. Research conducted by Chang et al. (Reference Chang, Womer, Edmiston, Bai, Zhou, Jiang and Wang2018) also reported that BD, MDD, and SZ patients presented gray matter volume decreases in 87.9% of the total regional volume with significant variations in the dACC, subgenual ACC, posterior cingulate cortex (PCC), temporal pole, insula, parahippocampus, angular gyri, cuneus, orbital frontal, and dorsal lateral prefrontal cortices. In more recent years, a similar trend has emerged from multisite mega-analyses conducted by the Enhancing Neuro Imaging Genetics Through Meta Analysis (ENIGMA) consortium. Specifically, Okada et al. (Reference Okada, Fukunaga, Miura, Nemoto, Matsumoto, Hashimoto and Hashimoto2023) found a consistent reduction in hippocampal volume in individuals with BD and SZ, whereas Opel et al. (Reference Opel, Goltermann, Hermesdorf, Berger, Baune and Dannlowski2020) demonstrated shared cortical abnormalities in BD, MDD, SZ, and obsessive-compulsive disorder, especially in the fusiform gyrus, hippocampus, and AI.
The potential existence and regional distribution of disorder-selective neuroanatomical variations that are associated with psychiatric conditions remain elusive. Tackling this deficiency in knowledge is important for multiple factors. First, uncovering disorder-selective neuroanatomical signatures could significantly support current endeavors in identifying robust neural biomarkers for differential diagnoses or measuring treatment outcomes in psychiatric disorders. Second, prioritizing research on targeted neural populations can facilitate the elucidation of the etiological mechanisms underlying clear-cut diagnostic categories, paving the way for targeted interventions, such as noninvasive brain stimulation. Third, addressing this translational challenge may refine current psychiatric nosology and ultimately inform future classifications and clinical assessments.
A secondary-level examination of the neuroanatomical variations tied to psychiatric disorders across the entire brain may be essential for answering this question. Specifically, 25 years of voxel-based morphometry (VBM) analysis (Ashburner & Friston, Reference Ashburner and Friston2000) of sMRI data in clinical research, along with the establishment of accessible and largely automated neuroimaging data repositories such as BrainMap (Vanasse et al., Reference Vanasse, Fox, Barron, Robertson, Eickhoff, Lancaster and Fox2018), provide an unparalleled resource for the systematic exploration of potential diagnosis-selective neural profiles for prominent psychiatric disorders. In the present study, we conducted a meta-analytic investigation of structural brain abnormalities across the whole brain, focusing on regional gray matter reductions derived from published VBM experiments using a traditional case–control design and analyzing 98 different diagnostic categories (33 psychiatric disorders and 65 neurological diseases). To achieve this goal, we introduce the Bayes fACtor mOdeliNg (BACON) approach (Costa et al., Reference Costa, Manuello, Ferraro, Liloia, Nani, Fox and Cauda2021), which integrates data from both mega- and meta-analyses to enable a quantitative, voxelwise characterization of diagnosis-selective neuroanatomical phenotypes in patients with BD, MDD, and SZ. Unlike conventional neuroimaging approaches using disorder-to-alteration mapping (Liloia, Costa, Cauda, & Manuello, Reference Liloia, Costa, Cauda and Manuello2024), BACON provides an alteration-to-disorder framework, quantifying the posterior probability that an observed structural variation is selectively associated with a target diagnosis relative to variations reported across other clinically defined categories.
Methods and materials
Data search
Adhering to PRISMA (Page et al., Reference Page, McKenzie, Bossuyt, Boutron, Hoffmann, Mulrow and Moher2021) and neuroimaging meta-analysis (Müller et al., Reference Müller, Cieslik, Laird, Fox, Radua, Mataix-Cols and Eickhoff2018) guidelines, we identified independent VBM experiments contrasting gray matter in healthy controls with patients diagnosed with BD, MDD, and SZ. Additionally, VBM experiments involving comparisons of healthy controls with other psychiatric disorders or neurological diseases stored in the BrainMap database (Vanasse et al., Reference Vanasse, Fox, Barron, Robertson, Eickhoff, Lancaster and Fox2018) were included. Only coordinates reporting gray matter volume/concentration decreases (i.e. healthy controls > disorders of interest) in a standardized stereotactic space were selected. To increase spatial accuracy, analyses were conducted in Talairach space with Montreal Neurological Institute (MNI) coordinates converted to Talairach space using the icbm2tal algorithm (Laird et al., Reference Laird, Robinson, McMillan, Tordesillas-Gutiérrez, Moran, Gonzales and Lancaster2010). A detailed description of eligibility criteria and search methodology is provided in the eMethods section in the Supplement. Since the study utilized published peer-reviewed data, institutional review board approval and patient consent were not required.
Database construction
For each experiment, coordinates of gray matter variations derived exclusively from whole-brain analyses were extracted. Coordinates based on region-of-interest or small-volume correction VBM analyses were excluded to avoid artificial biases favoring specific brain regions (Müller et al., Reference Müller, Cieslik, Laird, Fox, Radua, Mataix-Cols and Eickhoff2018). Using the CBMAT toolbox (Manuello et al., Reference Manuello, Liloia, Crocetta, Cauda and Costa2023), gray matter coordinates located in the cerebellum or outside the gray matter mask were excluded, as the cerebellum is frequently omitted in MRI scans, potentially introducing unreliability (Van Overwalle et al., Reference Van Overwalle, Ma, Haihambo, Bylemans, Catoira, Firouzi and Deroost2024). This exclusion also mitigates spurious clusters of selective variation in white matter structures.
These steps allowed us to define five data groupings essential for subsequent analyses: (1) the BD dataset, composed of coordinates describing gray matter variations in patients with BD (versus healthy control subjects); (2) the MDD dataset, with coordinates of gray matter variations in patients with MDD (versus healthy control); (3) the SZ dataset, with coordinates of gray matter variations in patients with SZ (versus healthy control); (4) the BrainMap psychiatric dataset, composed of coordinates describing gray matter variations in patients with other psychiatric disorders (versus healthy control subjects), sourced from the BrainMap database; and (5) the BrainMap psychiatric and neurological dataset, composed of coordinates describing gray matter variations in patients with other psychiatric disorders and neurological diseases (versus healthy control subjects) sourced from the BrainMap database. Further details of the database construction are provided in Figure 1A.
Overview of the analytical procedures. (A) A total of 8,740 coordinates of significant gray matter variation from 29,540 patients were extracted from 1,021 published voxel-based morphometry experiments. (B) Graphical representation of the data analytic pipeline, from the variation coordinates to the selectivity whole-brain map. The final statistical parametric map, which represents the values of the selective probability of the disorder of interest, is obtained with the Bayes factor computation implemented in the Bayes fACtor mOdeliNg plugin. (C) Data groupings for estimating the selectivity of the gray matter landscape in BD (analysis 1), MDD (analysis 2), and SZ (analysis 3) patients. (D) Robustness and supplementary functional and behavioral analyses of the Bayes fACtor mOdeliNg results. Note: BACON = Bayes fACtor mOdeliNg; BD = bipolar disorder; BF = Bayes factor; EXP. = experiment; MDD = major depressive disorder; SZ = schizophrenia.
Bayes fACtor mOdeliNg
We applied the BACON algorithm (Costa et al., Reference Costa, Manuello, Ferraro, Liloia, Nani, Fox and Cauda2021) to estimate posterior probabilities for gray matter variations selectively associated with BD, MDD, and SZ. This Bayesian-based approach evaluates the likelihood that observed neuroanatomical variations correspond to a disorder of interest.
We first conducted two meta-analyses via the anatomical likelihood estimation (ALE) method (Eickhoff et al., Reference Eickhoff, Bzdok, Laird, Kurth and Fox2012) for each disorder of interest: the first based on the coordinates of the disorder under investigation (e.g. the SZ dataset), and the second based on the coordinates of everything but the disorder under investigation coordinates (e.g. the merging of the BD, MDD, and BrainMap psychiatric datasets) (Figure 1B). The ALE algorithm, as implemented in the GingerALE software (v.3.0.2), was employed. The ALE method models the coordinates from each experiment as a series of three-dimensional Gaussian distributions of probabilities centered on the peaks of variations reported by the included experiments (Eickhoff et al., Reference Eickhoff, Nichols, Laird, Hoffstaedter, Amunts, Fox and Eickhoff2016). This permits the generation of a modeled alteration map for each experiment included in the meta-analysis. The size of the Gaussian kernel varies between the modeled alteration maps, taking into account the sample size originally used in each experiment. The union of all modeled alteration maps produces whole-brain voxelwise ALE scores, which quantify the degree of overlap among reported results at specific brain locations (Eickhoff et al., Reference Eickhoff, Bzdok, Laird, Kurth and Fox2012) (Figure 1B). A comprehensive statistical explanation of the method can be found in the eMethods section of the Supplement.
Second, we utilized the BACON algorithm as implemented in the MANGO software (v.4.1). BACON combines Bayes factor (Kass & Raftery, Reference Kass and Raftery1995) analysis with ALE-derived maps to determine the likelihood that observed neuroanatomical variations in a particular voxel are selectively associated with the disorder of interest rather than with other conditions (e.g. SZ versus no-SZ). This approach allowed us to evaluate two competing hypotheses in a whole-brain voxelwise manner: one positing that the variation is most likely related to the disorder under investigation and another suggesting that it is also associated with other conditions. In the absence of knowledge about the prior probabilities of these hypotheses, they are treated as equally likely, a choice supported by previous validation studies (Cauda et al., Reference Cauda, Nani, Liloia, Manuello, Premi, Duca and Costa2020; Costa et al., Reference Costa, Manuello, Ferraro, Liloia, Nani, Fox and Cauda2021; Liloia et al., Reference Liloia, Cauda, Uddin, Manuello, Mancuso, Keller and Costa2023). This assumption ensures that the computed Bayes factor directly reflects the relative strength of evidence for the disorder of interest compared with alternative clinical conditions. Thus, BACON computes posterior probabilities to quantify the likelihood that the gray matter variation observed in a voxel is selectively associated with the disorder under investigation, representing the probability P (disorder of interest | variation). A detailed statistical explanation is provided in the eMethods section of the Supplement, along with a visual illustration in Figure 1B.
Main analyses (psychiatric disorders)
We conducted three primary analyses (Figure 1C) to identify clusters showing diagnosis-selective structural variation for BD, MDD, and SZ relative to other psychiatric disorders in the BrainMap database. Results were thresholded at P (disorder of interest | variation)
$ \ge $
0.95 (i.e. a posterior probability of diagnosis-selectivity ≥ 0.95 for the target disorder) with a minimum cluster size of 500 mm3. This posterior-probability threshold is interpreted on the Bayes factor scale as ‘strong evidence’, following the classification of evidence strength by Kass and Raftery (Reference Kass and Raftery1995).
Additional analyses (psychiatric disorders and neurological diseases)
Three secondary analyses were also conducted for BD, MDD, and SZ patients, specifically combining psychiatric disorders and neurological diseases within the non-interest dataset. Consistent with the primary analyses, results were thresholded at P (disorder of interest | variation)
$ \ge $
0.95.
Robustness analyses
Meta-analytic findings may be influenced by the file-drawer problem, a form of publication bias where experiments with null or contradictory results remain unpublished. To address this issue, we assessed the robustness of our findings via the fail-safe technique adapted for the neuroimaging meta-analytic environment (Acar, Seurinck, Eickhoff, & Moerkerke, Reference Acar, Seurinck, Eickhoff and Moerkerke2018). Based on a recent simulation study (Samartsidis et al., Reference Samartsidis, Montagna, Laird, Fox, Johnson and Nichols2020), estimating a 6% rate of missing experiments in the BrainMap database, we retested our analyses by introducing an equivalent percentage of noise (i.e. random simulated coordinates of variation) to our datasets of no interest (Figure 1D). Robustness was further evaluated by increasing noise levels up to 30%, in line with recent recommendations (Gray et al., Reference Gray, Müller, Eickhoff and Fox2020). Details are provided in the eMethods in the Supplement.
Functional and behavioral analyses
Supplementary post hoc analyses were conducted to further characterize and interpret the robust meta-analytic cluster identified via BACON. These analyses encompassed task-based coactivation characterization via meta-analytic connectivity modeling (MACM) (Laird et al., Reference Laird, Eickhoff, Rottschy, Bzdok, Ray and Fox2013) and observer-independent brain-to-behavior association via the Behavioral plugin (Lancaster et al., Reference Lancaster, Laird, Eickhoff, Martinez, Fox and Fox2012) (Figure 1D). The analyses were performed using the BrainMap functional database (Fox & Lancaster, Reference Fox and Lancaster2002; Laird et al., Reference Laird, Eickhoff, Kurth, Fox, Uecker, Turner and Fox2009), which focuses on functional MRI experiments of healthy participants involved in normal mapping task-based experiments (8,377 eligible experiments). MACM was determined via ALE-based methods with a cluster-level familywise error correction threshold of P < 0.05 and a cluster-forming threshold of P < 0.001 (Eickhoff et al., Reference Eickhoff, Nichols, Laird, Hoffstaedter, Amunts, Fox and Eickhoff2016). A Bonferroni-corrected threshold at P < 0.05 was adopted to designate statistically significant behavioral associations (Lancaster et al., Reference Lancaster, Laird, Eickhoff, Martinez, Fox and Fox2012). A detailed description of these analyses can be found in the eMethods section in the Supplement.
Results
A total of 1,021 individual neuroimaging experiments, comprising 29,540 patients and 28,177 healthy controls, were included in the study (Figure 2). The included diagnostic groups were BD (73 VBM experiments; Supplementary Table S1), MDD (82 VBM experiments; Supplementary Table S2), SZ (123 VBM experiments; Supplementary Table S3), other psychiatric disorders from the BrainMap database (30 different disorders and 175 VBM experiments; Supplementary Table S4), and other psychiatric disorders combined with neurological diseases from the BrainMap database (95 diagnostic categories and 743 VBM experiments; Supplementary Table S5). Further details of these groups are provided in Figure 1A, Supplementary Tables S6 and S7.
Overview of literature selection and coding (PRISMA flowchart). Note: BD = bipolar disorder; MDD = major depressive disorder; N = number of; ROI = region-of-interest; SZ = schizophrenia disorder; SVC = small volume correction; WM = white matter; VBM = voxel-based morphometry.
Selective variation profile of bipolar disorder
The primary BACON analysis using psychiatric data (Supplementary Tables S4 and S6) revealed three clusters of variation at P
$ \ge $
0.95. Specifically, we observed gray matter selectivity for BD patients in the right middle frontal gyrus, right middle temporal gyrus (MTG), and left inferior parietal lobule (Table 1A). However, fail-safe analysis revealed that only the cluster in the right MTG was robust (Figure 3A), withstanding simulated data injection by 30% of experiments, whereas the others were negated with just 6% of injected simulated coordinates, which is consistent with the estimated proportion of missing experiments in the BrainMap database (Samartsidis et al., Reference Samartsidis, Montagna, Laird, Fox, Johnson and Nichols2020).
Clusters of selective gray matter variation in bipolar disorder (A), major depressive disorder (B), and schizophrenia (C) derived from Bayes fACtor mOdeliNg analysis of psychiatric-only data and thresholded at P (disorder-of-interest | variation)
$ \ge $
0.95 (95%)
Abbreviations: BA = Brodmann area; BACON = Bayes fACtor mOdeliNg.
a Robust clusters of variation selectivity, remaining stable over 30% of injected simulated experiments.
Robust clusters of selective gray matter variation in bipolar disorder and schizophrenia and their network-level functional and behavioral characterizations. (A) Brain cluster of selective gray matter variation in bipolar disorder derived from Bayes fACtor mOdeliNg analysis of psychiatric disorders data thresholded at P (bipolar disorder | variation)
$ \ge $
0.95 (95%) and remaining stable over 30% of injected random experiments via fail-safe analysis. On the right side, the meta-analytic coactivation map (MACM) shows areas coactivated with the cluster of variation in healthy participant task-based activation experiments in the BrainMap database, as well as its statistically linked physiological mental processes derived from the behavioral analysis of the BrainMap database. (B) Brain cluster of selective gray matter variation in schizophrenia derived from Bayes fACtor mOdeliNg analysis of psychiatric disorders data thresholded at P (schizophrenia | variation)
$ \ge $
0.95 (95%) and remaining stable over 30% of injected random experiments via fail-safe analysis. On the right side, the MACM shows areas coactivated with the cluster of variation in healthy participant task-based activation experiments in the BrainMap database, as well as its statistically linked physiological mental processes derived from the behavioral analysis of the BrainMap database. The Bayes fACtor mOdeliNg results are visualized as hemispheric surfaces (three-dimensional view). MACM results are visualized as axial slices (two-dimensional cortical and subcortical views). Templates are in neurological convention. Note: ACC = posterior dorsal anterior cingulate cortex; ALE = activation likelihood estimation; MTG = middle temporal gyrus.
To determine whether this selective cluster is normally part of a coherent functional circuit, we examined its pattern of task-dependent coactivation. MACM analysis revealed significant coactivation with the left middle and superior temporal gyri (Brodmann areas, BAs 39), as well as the bilateral PCC (BAs 31), and the inferior temporal gyri (BAs 21) (Figure 3A and Supplementary Table S8). Thus, we used the Behavioral plugin to evaluate the corresponding physiological mental processes statistically associated with the network of coactivation. This resulted in the identification of six subdomains, which reflected the cognitive functioning domain (i.e. social cognition, explicit memory, language semantics, reasoning, and language syntax). Further details are provided in Figure 3A and Supplementary Table S9.
In contrast, the additional BACON analysis of psychiatric and neurological conditions did not reveal any selective gray matter clusters for BD at P
$ \ge $
0.95.
Selective variation profile of major depressive disorder
The primary BACON analysis and additional analysis revealed no selective gray matter clusters for MDD at thresholds of P
$ \ge $
0.95.
Selective variation profile of schizophrenia
The primary BACON analysis revealed two clusters of variation at P
$ \ge $
0.95. Specifically, we observed gray matter selectivity for SZ patients in the right posterior dACC (also known as caudal ACC) (Table 1C). Fail-safe analysis confirmed the robustness of the right posterior dACC cluster, which remained stable across 30% of simulated coordinates. However, the other cluster was negated with 6% of simulated data injections (Figure 3B).
MACM demonstrated significant coactivation of the right posterior dACC with several regions, including the left ACC, bilateral middle frontal gyri (BAs 6), AI (BA 13), inferior frontal gyri (BAs 9 and 44), precuneus (BA 7), fusiform gyri (BAs 19 and 37), putamen, globus pallidum, and thalamus (Figure 3B and Supplementary Table S10). The corresponding physiological mental processes statistically associated with the right posterior dACC network of coactivation were 49, which reflected all functioning domains of the BrainMap database (i.e. action, cognition, emotion, interoception, and perception). Details can be found in Figure 3B and Supplementary Table S11.
The additional BACON analysis of psychiatric and neurological conditions did not identify any selective gray matter clusters for SZ at P
$ \ge $
0.95.
Discussion
Leveraging 25 years of peer-reviewed VBM clinical research, this study yielded several important findings. We provide strong evidence of robust selectivity in the neuroanatomical phenotypes associated with BD and SZ, compared to other psychiatric conditions. Furthermore, we reveal that selective variation appears to accumulate in well-delineated high-order areas, which display extensive coactivation across large-scale networks and are consistently engaged in cognitive domains impaired in these disorders.
Building on the foundational work of Goodkind et al. (Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata and Etkin2015), extensive brain mapping research has identified a shared neuroanatomical substrate across mental illnesses (Hettwer et al., Reference Hettwer, Larivière, Park, van den Heuvel, Schmaal, Andreassen and Valk2022; Kaczkurkin et al., Reference Kaczkurkin, Park, Sotiras, Moore, Calkins, Cieslak and Satterthwaite2019; Okada et al., Reference Okada, Fukunaga, Miura, Nemoto, Matsumoto, Hashimoto and Hashimoto2023; Opel et al., Reference Opel, Goltermann, Hermesdorf, Berger, Baune and Dannlowski2020; Romer et al., Reference Romer, Elliott, Knodt, Sison, Ireland, Houts and Hariri2021; Wise et al., Reference Wise, Radua, Via, Cardoner, Abe, Adams and Arnone2017). In this study, we advanced this line of inquiry by adopting a data-driven, hypothesis-free approach to examine whether BD, MDD, and SZ are distinguished by selective structural brain phenotypes. Our primary analyses identified localized multimodal regions in the right hemisphere for both BD and SZ. However, other regions considered neuroanatomical markers of these disorders (i.e. amygdala, basal ganglia, dorsal and subgenual ACC, fusiform gyrus, hippocampus, inferior frontal gyrus, insula, prefrontal cortex, temporal pole, and thalamus) (Chen et al., Reference Chen, Wang, Gong, Qi, Fu, Tang and Wang2022; Liloia et al., Reference Liloia, Brasso, Cauda, Mancuso, Nani, Manuello and Rocca2021) exhibited no significant selectivity, indicating their shared involvement across multiple psychiatric categories. These findings expand prior transdiagnostic research by integrating whole-brain cross-disorder similarity in morphometry-derived phenotypes with disorder-selective loci previously uncharacterized.
We found robust clusters of selective variation in the right MTG and posterior dACC for BD and SZ, respectively. These findings are particularly significant given prior VBM meta-analyses (Chen et al., Reference Chen, Wang, Gong, Qi, Fu, Tang and Wang2022; Liloia et al., Reference Liloia, Brasso, Cauda, Mancuso, Nani, Manuello and Rocca2021; Wise et al., Reference Wise, Radua, Via, Cardoner, Abe, Adams and Arnone2017; Zhu et al., Reference Zhu, Wang, Zhou, Fang, Huang, Xie and Chen2022) that have consistently reported morphometric aberrations in these regions across primary investigations. This reinforces their potential role in current neurobiological models of BD and SZ. The MTG has been repeatedly mentioned in neuroimaging research of BD, showing cortical thickness and functional reductions (Hibar et al., Reference Hibar, Westlye, Doan, Jahanshad, Cheung, Ching and Andreassen2018; Kuang et al., Reference Kuang, Gao, Long, Cao, Cui, Guo and Lu2022; Long et al., Reference Long, Wang, Tian, Cao, Xie and Jia2022), along with disruptions in both long- and short-range connectivity (Liu et al., Reference Liu, Jiang, Deng, Jia, Sun, Kong and Tang2022; Syan et al., Reference Syan, Minuzzi, Smith, Allega, Hall and Frey2017; Wang et al., Reference Wang, Zhong, Jia, Sun, Wang, Liu and Huang2016). In line with our observer-independent behavioral analyses, this area has been shown to support key functions aberrant in the disorder, such as social cognition, reasoning, semantic processing, and memory (Keramatian, Torres, & Yatham, Reference Keramatian, Torres and Yatham2021). Therefore, these findings offer a new perspective on the importance of this region, which may provide critical insights into core cognitive impairments in BD. Future multidisciplinary diagnostic research integrating the right MTG is warranted.
Similarly, the selective involvement of the right posterior dACC represents a notable finding for translational SZ research. The ACC has been widely recognized as a key morpho-functional hub in SZ (Baiano et al., Reference Baiano, David, Versace, Churchill, Balestrieri and Brambilla2007; Vitolo et al., Reference Vitolo, Tatu, Pignolo, Cauda, Costa, Ando’ and Zennaro2017; Wang et al., Reference Wang, Zhou, Zhuo, Qin, Zhu, Liu and Yu2015) and identified as a putative epicenter of neural tissue loss (Shafiei et al., Reference Shafiei, Markello, Makowski, Talpalaru, Kirschner, Devenyi and Mišić2020). However, its extensive involvement across various mental illnesses (Cauda et al., Reference Cauda, Nani, Manuello, Liloia, Tatu, Vercelli and Costa2019; Chang et al., Reference Chang, Womer, Edmiston, Bai, Zhou, Jiang and Wang2018; Goodkind et al., Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata and Etkin2015; Sha, Wager, Mechelli, & He, Reference Sha, Wager, Mechelli and He2019; Wise et al., Reference Wise, Radua, Via, Cardoner, Abe, Adams and Arnone2017) renders our findings unique. Importantly, the dACC cluster identified in our study does not overlap with the previously reported gray matter loss in the ACC observed across psychiatric categories in transdiagnostic studies such as Goodkind et al. (Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata and Etkin2015). Instead, our findings suggest that the more caudal portions of the dACC may have a central role in SZ, in contrast to the dorsal anterior and subgenual regions commonly implicated across psychiatric disorders (Chang et al., Reference Chang, Womer, Edmiston, Bai, Zhou, Jiang and Wang2018; Goodkind et al., Reference Goodkind, Eickhoff, Oathes, Jiang, Chang, Jones-Hagata and Etkin2015; Wise et al., Reference Wise, Radua, Via, Cardoner, Abe, Adams and Arnone2017). Our results are consistent with the ‘schizophrenia neural common network’ described via coordinate network mapping (CNM) (Makhlouf et al., Reference Makhlouf, Drew, Stubbs, Taylor, Liloia, Grafman and Siddiqi2025), a multimodal approach incorporating a network-level design to address inter-subject heterogeneity in VBM studies. Similarly, they align with a recent mega-analysis from the ENIGMA consortium, which identified low structural heterogeneity in cortical folding of the right caudal ACC across a multi-site cohort of 5,626 subjects with SZ (Omlor et al., Reference Omlor, Rabe, Fuchs, Surbeck, Cecere, Huang and Homan2025). Although our BACON-derived cluster was more sensitive than the CNM-defined and ROI-based cortical findings, this concordance emphasizes the value of integrating diverse neuroimaging methodologies to enhance our understanding of the neurobiological foundations of SZ, facilitating the development of targeted interventions.
No clusters of variation were found in MDD. This result is consistent with the lack of spatial consistency in gray matter reduction identified by recent VBM meta-analyses (Gray et al., Reference Gray, Müller, Eickhoff and Fox2020; Müller et al., Reference Müller, Cieslik, Serbanescu, Laird, Fox and Eickhoff2017), which cannot be attributed to insufficient statistical power. Importantly, multisite mega-analysis studies from the ENIGMA consortium (Cheon et al., Reference Cheon, Bearden, Sun, Ching, Andreassen, Schmaal and van Erp2022; Hibar et al., Reference Hibar, Westlye, Doan, Jahanshad, Cheung, Ching and Andreassen2018, Reference Hibar, Westlye, van Erp, Rasmussen, Leonardo, Faskowitz and Andreassen2016; Schmaal et al., Reference Schmaal, Hibar, Sämann, Hall, Baune, Jahanshad and Veltman2017, Reference Schmaal, Pozzi, Ho, van Velzen, Veer, Opel and Veltman2020; van Erp et al., Reference van Erp, Hibar, Rasmussen, Glahn, Pearlson, Andreassen and Turner2016, Reference van Erp, Walton, Hibar, Schmaal, Jiang, Glahn and Turner2018) further suggest an ‘affective-psychotic severity brain continuum’, with widespread cortical and subcortical abnormalities in SZ, intermediate involvement in BD, and more localized, less pronounced variations in MDD. In parallel, differences in average clinical burden across these disorders, including greater functional impairment and a higher burden of severe psychotic symptoms in SZ, intermediate levels in BD, and psychotic features in MDD largely confined to specific subtypes (Aminoff et al., Reference Aminoff, Onyeka, Ødegaard, Simonsen, Lagerberg, Andreassen and Melle2022; Bowie et al., Reference Bowie, Depp, McGrath, Wolyniec, Mausbach, Thornquist and Pulver2010; Jääskeläinen et al., Reference Jääskeläinen, Juola, Korpela, Lehtiniemi, Nietola, Korkeila and Miettunen2018; Ohayon & Schatzberg, Reference Ohayon and Schatzberg2002), may plausibly be associated with more spatially consistent neuroanatomical alterations in SZ and, to a lesser extent, BD relative to MDD. Furthering this narrative, MDD represents a heterogeneous disorder that, according to the DSM-5 or ICD-11 criteria, may be diagnosed with more than 200 different combinations of signs and symptoms pooling together, within the same nosographic category. These patients can exhibit completely different sets of psychopathological manifestations (Zimmerman et al., Reference Zimmerman, Ellison, Young, Chelminski and Dalrymple2015). This phenotypic heterogeneity corresponds to a marked diversity in genetic and environmental risk factors, which, interacting with each other in a multitude of different ways, could underlie the observed variability at the neural level (Müller et al., Reference Müller, Cieslik, Serbanescu, Laird, Fox and Eickhoff2017; Nguyen et al., Reference Nguyen, Harder, Xiong, Kowalec, Hägg, Cai and Lu2022; Zhang et al., Reference Zhang, Sweeney, Bishop, Gong and Lui2023).
The additional analysis of BD, MDD, and SZ data against 65 neurological diseases revealed that regional reductions identified via the VBM technique exhibit a distributed, nonselective pattern crossing psychiatric and neurological boundaries. This finding corroborates previous studies (Cauda et al., Reference Cauda, Nani, Manuello, Liloia, Tatu, Vercelli and Costa2019, Reference Cauda, Nani, Manuello, Premi, Palermo, Tatu and Costa2018; Crossley et al., Reference Crossley, Mechelli, Scott, Carletti, Fox, McGuire and Bullmore2014; de Lange et al., Reference de Lange, Scholtens, van den Berg, Boks, Bozzali, Cahn and van den Heuvel2019; Vanasse et al., Reference Vanasse, Fox, Fox, Cauda, Costa, Smith and Lancaster2021) demonstrating a shared neural architecture across psychiatric and neurological conditions. Moreover, pathophysiological commonalities have been consistently observed at multiple biological levels, including genetic and molecular evidence (Bryois et al., Reference Bryois, Calini, Macnair, Foo, Urich, Ortmann and Malhotra2022; Cauda et al., Reference Cauda, Nani, Manuello, Premi, Palermo, Tatu and Costa2018; Peall, Owen, & Hall, Reference Peall, Owen and Hall2024; Smeland et al., Reference Smeland, Kutrolli, Bahrami, Fominykh, Parker, Fuhrer, Hindley, Rødevand, Jaholkowski, Tesfaye, Parekh, Elvsåshagen, Grotzinger, Steen, van der Meer, O’Connell, Djurovic, Dale, Shadrin and Andreassen2025). Our findings are also aligned with the cross-disorder dysconnectivity hypothesis of the human connectome (van den Heuvel & Sporns, Reference van den Heuvel and Sporns2019). This is particularly evident in neurodegenerative diseases, where processes such as the transneuronal transportation of toxic misfolded proteins, trophic failure, and nodal stress degeneration drive neural death and atrophy (Raj & Powell, Reference Raj and Powell2018). These processes result in network-like degeneration patterns that closely mirror the structural, functional, and genetic architecture of brain connectivity (Cauda et al., Reference Cauda, Nani, Manuello, Premi, Palermo, Tatu and Costa2018; Fornito, Zalesky, & Breakspear, Reference Fornito, Zalesky and Breakspear2015; Raj & Powell, Reference Raj and Powell2018). Importantly, over the disease course, these degeneration patterns gradually spread throughout the brain (Wilson et al., Reference Wilson, Cookson, Van Den Bosch, Zetterberg, Holtzman and Dewachter2023), suggesting the possibility that widespread neural disruption encompasses regions implicated in psychiatric disorders. This raises the intriguing hypothesis that the observed degeneration patterns might reflect a unifying neurobiological substrate, potentially linking psychiatric and neurological conditions within a common pathophysiological framework (Cauda et al., Reference Cauda, Nani, Manuello, Liloia, Tatu, Vercelli and Costa2019; Crossley et al., Reference Crossley, Mechelli, Scott, Carletti, Fox, McGuire and Bullmore2014; de Lange et al., Reference de Lange, Scholtens, van den Berg, Boks, Bozzali, Cahn and van den Heuvel2019).
From a neurobiological perspective, do the present results support a categorical or a neural general dimension of psychopathology? We have demonstrated that localized, diagnosis-selective variations can be associated with BD and SZ. Nevertheless, the low probability of selectivity we identified across the rest of the brain is also reminiscent of recent dimensional models, such as the general psychopathology factor (P-factor) (Caspi et al., Reference Caspi, Houts, Belsky, Goldman-Mellor, Harrington, Israel and Moffitt2014). This model posits that the meta-structure of mental illnesses and their neural equivalent can be understood hierarchically, comprising both a single general dimension and several specific dimensions at the lower levels of the hierarchy (Caspi & Moffitt, Reference Caspi and Moffitt2018; Sprooten, Franke, & Greven, Reference Sprooten, Franke and Greven2022). While this study provides potential neural insights that may contribute to the interpretation of a ‘neural P-factor’, the validity and substantive meaning of this dimension remain subjects of ongoing debate (DeYoung et al., Reference DeYoung, Kotov, Krueger, Cicero, Conway, Eaton and Wright2022; Haeffel et al., Reference Haeffel, Jeronimus, Kaiser, Weaver, Soyster, Fisher and Lu2022; Sprooten et al., Reference Sprooten, Franke and Greven2022). Finally, it is necessary to note that the underlying cellular mechanisms responsible for changes in VBM signals in mental illnesses remain elusive (Mancuso et al., Reference Mancuso, Fornito, Costa, Ficco, Liloia, Manuello and Cauda2020). Understanding the pathophysiological principles driving neuroanatomical variations is a crucial next step for unraveling the biological basis of disorder variability and selectivity. We aspire that the presented meta-analytic features shall facilitate and provide a novel coordinate frame for such translational insights.
Methodological considerations
The current meta-analytic study proposes a novel and statistically rigorous perspective on the complex neuroanatomical architecture underlying three prominent psychiatric disorders. In comparison to prior qualitative reviews and meta-analyses, our methodological design advances the field on several levels. We assembled the largest VBM meta-analytic dataset to date, examining data from 98 distinct diagnostic categories. This approach allowed us to optimize the balance between sensitivity and susceptibility to false-positive effects, thereby enhancing the statistical power of our quantitative synthesis (Manuello, Costa, Cauda, & Liloia, Reference Manuello, Costa, Cauda and Liloia2022; Müller et al., Reference Müller, Cieslik, Laird, Fox, Radua, Mataix-Cols and Eickhoff2018). Also, the innovative transdiagnostic approach allowed us to extend the clinical information given by canonical neuroimaging meta-analyses focusing on a single disorder of interest, providing valuable insights into the selective role of each aberrant neuroanatomical component of clinical conditions under investigation (Cauda et al., Reference Cauda, Nani, Liloia, Manuello, Premi, Duca and Costa2020; Liloia et al., Reference Liloia, Costa, Cauda and Manuello2024). Not less important, by employing a Bayesian-based framework, we moved beyond binary frequentist concepts of statistical significance (i.e. dichotomous reject/do-not-reject framework), providing a direct probabilistic hypothesis assessment (Costa et al., Reference Costa, Manuello, Ferraro, Liloia, Nani, Fox and Cauda2021; Friston et al., Reference Friston, Penny, Phillips, Kiebel, Hinton and Ashburner2002). Finally, although this study proposed a new outlook on the selective brain architectures of BD, MDD, and SZ, it is important to note that the BACON methodology can potentially be applied to any other clinical condition reporting neuroanatomical variations, opening attractive translational prospects for an in-depth comprehension of the clinical brain.
Despite these strengths, several limitations warrant consideration. By definition, neuroimaging meta-analyses are associated with lower spatial resolution than native statistical parametric maps, as they rely on reported stereotactic coordinates rather than full voxelwise data (Müller et al., Reference Müller, Cieslik, Laird, Fox, Radua, Mataix-Cols and Eickhoff2018). Although our experimental design employed the anatomical ALE, the most widely used meta-analytic algorithm worldwide with demonstrated unbiased spatial reconstruction (Eickhoff et al., Reference Eickhoff, Nichols, Laird, Hoffstaedter, Amunts, Fox and Eickhoff2016; Manuello et al., Reference Manuello, Costa, Cauda and Liloia2022; Müller et al., Reference Müller, Cieslik, Laird, Fox, Radua, Mataix-Cols and Eickhoff2018; Radua & Mataix-Cols, Reference Radua and Mataix-Cols2012; Tahmasian et al., Reference Tahmasian, Sepehry, Samea, Khodadadifar, Soltaninejad, Javaheripour and Eickhoff2019), this constraint remains inherent. At the same time, it is important to note that in the absence of a publicly available repository of peer-reviewed voxelwise whole-brain MRI data, using coordinate-based databases such as BrainMap remains the unique approach to exploring disorder-selective brain variations. Nevertheless, coordinate-based inference is also contingent on the reporting practices of published VBM studies (Manuello et al., Reference Manuello, Costa, Cauda and Liloia2022). In particular, the historically incomplete reporting of GM increases in the psychiatric primary literature can affect meta-analytic synthesis of ‘healthy controls < disorders of interest’ contrasts (Mancuso et al., Reference Mancuso, Fornito, Costa, Ficco, Liloia, Manuello and Cauda2020), potentially introducing reporting-related bias and undermining interpretability when increases are the target of inference. For this reason, we did not include a dedicated case–control analysis focused on GM increases in the present work. At the same time, systematically characterizing increase-selective GM effects remains an important target for future research. Still, the cross-sectional secondary-level design precludes analysis of the potential impact on findings of key sociodemographic and clinical variables (e.g. age, sex, education, illness duration, presence of psychotic features during mood episodes, and medication status). While we do not consider this a limitation per se, as the primary aim of a neuroimaging meta-analysis is to overcome sample heterogeneity and identify invariant findings across groups of interest (Eickhoff et al., Reference Eickhoff, Bzdok, Laird, Kurth and Fox2012; Fox, Lancaster, Laird, & Eickhoff, Reference Fox, Lancaster, Laird and Eickhoff2014; Manuello et al., Reference Manuello, Costa, Cauda and Liloia2022; Tahmasian et al., Reference Tahmasian, Sepehry, Samea, Khodadadifar, Soltaninejad, Javaheripour and Eickhoff2019), disregarding the unique characteristics of individual subjects or diagnostic subgroups may be overly simplistic and risk overlooking critical features. We propose that future research could develop subject-level implementations of the BACON methodology to assess the neuroanatomical selectivity of variations and explore whether this approach can discriminate between different categories or dimensions within the disorder of interest. Currently, the principal challenge in extending this methodology to subject-level data lies in identifying focal neuroanatomical variations in the absence of normative intensity values, which are necessary to reliably distinguish T1 images of healthy versus clinical subjects (Arbabshirani, Plis, Sui, & Calhoun, Reference Arbabshirani, Plis, Sui and Calhoun2017; Bzdok & Karrer, Reference Bzdok, Karrer, Diwadkar, Eickhoff and Di2021; Liloia et al., Reference Liloia, Brasso, Cauda, Mancuso, Nani, Manuello and Rocca2021; Scarpazza & Simone, Reference Scarpazza and Simone2016).
Conclusions
Our secondary level findings indicate that diagnosis-selective and robust neuroanatomical variations are identifiable in BD and SZ, but not in MDD. These results refine our understanding of the neuroanatomy of these complex disorders, opening attractive prospects for future neuroimaging-based translational research.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726103511.
Data availability statement
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
Acknowledgments
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. The data studied in this study are freely available as part of the BrainMap database (http://brainmap.org/). Data were analyzed using the GingerALE software package (version 3.0.2) that can be freely downloaded from https://www.brainmap.org/ale/. Data were analyzed using the Bayes factor modeling plug-in that can be freely downloaded from https://figshare.com/articles/software/Bacon_Plugin/12988661. Simulated data were generated using the Fail-Safe R script that can be freely downloaded from https://github.com/NeuroStat/GenerateNull. Data were analyzed using the Behavioral plug-in that can be freely downloaded from https://mangoviewer.com/plugin_behavioralanalysis.html.
Author contribution
DL: Conceptualization, methodology, formal analysis, investigation, software, validation, resources, data curation, writing – original draft, writing – review and editing, visualization. PR: Supervision, writing – review, and editing. CB: Writing – original draft, writing – review, and editing. MT: Writing – review and editing. JM: Formal analysis, resources, writing – review and editing. AC: Data curation, writing – review and editing. SD: Writing – review and editing. TC: Methodology, investigation, software, validation, resources, supervision, writing – review, and editing. FC: Methodology, investigation, software, validation, resources, data curation, supervision, writing – review and editing.
Competing interests
The authors declare none.
Open access
