Symptom-specific alterations in subregional intrinsic connectivity of anterior cingulate cortex in major depressive disorder Authors: Zilin Zhou, Lianqing Zhang, Xinyue Hu, Lingxiao Cao, Yingxue Gao, Hailong Li, Weijie Bao, Mengyue Tang, Huaiqiang Sun, Weihong Kuang, Graham J. Kemp, Xiaoqi Huang, Qiyong Gong Published: 2025 ABSTRACT The anterior cingulate cortex (ACC) is critically implicated in major depressive disorder (MDD), with distinct subregional dysfunctions related to specific aspects of emotional and cognitive dysregulation. However, its precise functional architecture, including the number, boundaries and intrinsic connectivity patterns of subdivisions, remains to be characterized in MDD. To address this, we employed data-driven connectivity-based parcellation (CBP) to delineate the optimal ACC functional partitions in 158 first-episode medication-naïve non-comorbid patients with MDD, and 152 healthy controls (HC). We generated subregion-specific ACC intrinsic connectivity maps, and analyzed group differences using diagnosis-by-subregion flexible factorial analyses of variance. We further explored the relationship between connectivity changes and clinical features. The ACC was optimally parcellated into bilateral ventral ACC (vACC) and dorsal ACC (dACC): The vACC showed preferential connections with default mode network (DMN) regions, whereas the dACC primarily connected to salience network and subcortical areas. Significant diagnosis-by-subregion interactions in ACC connectivity were mainly in the DMN regions: in HC, vACC showed stronger connectivity with DMN regions compared to the dACC, while this difference was attenuated in MDD, characterized by the vACC-DMN hypoconnectivity and a trend towards dACC-DMN hyperconnectivity. Clinically, lower vACC-DMN connectivity was related with core depressive symptoms, i.e., lower mood and longer illness duration, while higher dACC-DMN connectivity was linked to elevated maladaptive behaviors, e.g., suicide and insomnia. These findings confirm ventral-dorsal functional heterogeneity of ACC architecture and reveal symptom-specific patterns of subregional ACC-DMN dysconnectivity in MDD, which may provide insights for developing tailored interventions to address specific clinical manifestations. Introduction The anterior cingulate cortex (ACC) is a brain structure crucial for emotional processing, cognitive regulation, and homeostatic maintenance [1, 2]. Dysfunction of ACC has been strongly associated with stress-related negative emotions, anhedonia-like behaviors, ruminations, and cognitive deficits [3–6], which are commonly observed in major depressive disorder (MDD). Consequently, the ACC is increasingly recognized to play a pivotal role in the pathophysiology of MDD and the mechanisms underlying antidepressant treatments [7–9]. An early cellular-level study identified reduced neuronal soma size and enhanced neuronal density in cortical layers 5 and 6 of the ACC as key pathological features of MDD [10]. On the macrostructural level, meta-analyses and multicenter studies in MDD reported smaller gray matter volume and thinner cortical thickness in the ACC [11, 12]. Functional neuroimaging studies in MDD revealed abnormal ACC activation at rest and during tasks concerning negative emotion and requiring cognitive effort [13, 14]. Moreover, emerging evidence supported the view that MDD involves large-scale intrinsic dysconnectivity of ACC-centred networks [15, 16]. The ventral ACC, particularly its subgenual part, has emerged as the major target for network-based neuromodulation in depression [17–19], and its connectivity patterns have shown promise in predicting outcomes of antidepressant treatments [20, 21]. Dysfunctions in ACC-centered networks have been extensively revealed through seed-based functional connectivity analysis, a commonly used method for defining the functional architecture of spontaneously-coupled brain networks widely applied in psychiatric research [22]. Additionally, the ACC exhibits substantial functional heterogeneity along its ventral-dorsal partition [23, 24]. In healthy individuals, differentiated dorsoventral ACC subregions have distinct connectivity features: the dorsal seeds show preferential connectivity with frontoparietal and sensorimotor areas related to high-level cognitive functions, while the ventral seeds show consistent patterns of correlated activity with limbic regions and default mode network related to affective processes [25, 26]. A recent meta-analysis in MDD revealed divergent alterations in subregional ACC-centered intrinsic connectivity along its ventral-dorsal axis, possibly implicated in distinctive clinical manifestations, such as negative emotion bias, sensorimotor dysregulation, and cognitive impairment [16]. However, previous studies of subregional ACC connectivity alterations in MDD typically relied on geometric or anatomical partitioning based on predefined coordinates or atlases, and these partitions may not align well with the brain’s intrinsic functional organization. For example, a priori ACC subregional coordinates equally-spaced along the corpus callosum curvature yielded multiple subdivisions with substantial overlap in intrinsic connectivity profiles, especially between adjacent subregions [26]. Alternatively, the automated anatomical labelling (AAL3) atlas divided the ACC into three parts based on simple anatomical planes [27]. The general point is that any mismatch between predefined boundaries and true functional subregions would likely obscure the nuances of ACC subregional dysconnectivity [28, 29]. Furthermore, practical utility of literature-based definitions was limited by the cross-study variability in sample characteristics, scanning protocols, and preprocessing pipelines. To address these issues, we adopted a well-validated, data-driven connectivity-based parcellation (CBP) technique [30, 31] to delineate the sample-specific ACC functional architecture by leveraging intrinsic connectivity properties. Unlike geometric or anatomical parcellations, the CBP method optimizes resting-state functional homogeneity within subregions while preserving intrinsic connectivity architecture in specific population. It offers more precise and biologically meaningful functional subdivisions, facilitating the detection of subregional connectivity nuances. This approach has been successfully applied to cortical and subcortical partitioning in both healthy and clinical populations [30, 32]. Here, we used the CBP technique to identify optimal functional subdivisions of ACC, each characterized by distinct intrinsic connectivity patterns [30]. We further investigated subregion-specific alterations in ACC functional architecture and explored their clinical significance in a relatively large cohort of first-episode, medication-naïve, and non-comorbid patients with MDD. Based on previous findings [15, 16], we hypothesized the presence of two or three optimally functional subdivisions within the ACC along its ventral-dorsal partition, each demonstrating unique intrinsic connectivity profiles (hypothesis 1). Given the well-established functional distinction between ventral “affective” and dorsal “cognitive” parts of the ACC [23], we posited that ventral portions would exhibit dysconnectivity with emotion-related brain regions, while dorsal portions would show dysconnectivity with cognition-related brain regions in MDD (hypothesis 2). We further speculated that dysconnectivity of ventral ACC would correlate with mood disturbances, while that of dorsal ACC would associate with cognitive or behavioral problems (hypothesis 3). Materials and methods Participants Participants in this cross-sectional study were drawn from the Huaxi Depression Imaging Cohort Study (HX-DICS) at the West China Hospital of Sichuan University. Diagnosis of MDD was independently determined by two experienced psychiatrists based on the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID). The 17-item Hamilton Depression Scale (HAMD-17) and 14-item Hamilton Anxiety Scale (HAMA-14) were used to evaluate the severity of depression and anxiety, respectively. All subjects were right-handed according to the Edinburgh Handedness Inventory. Inclusion criteria for MDD patients were: met DSM-IV-TR diagnostic criteria for MDD; age 18-60 years; first-episode; had not received any psychiatric medication or any other interventions (e.g., psychotherapy or neuromodulation); total HAMD-17 score ≥ 17 (moderate and severe MDD) [33]; and not comorbid with any other Axis I psychiatric disorders. Exclusion criteria included: current or history of major physical illness, cardiovascular disease, neurological disorder, substance dependence or abuse; standard MRI scanning contraindications; pregnancy or lactation. A total of 158 first-episode, medication-naïve, and non-comorbid patients with MDD (59 males and 99 females, mean age 29.77 ± 9.82 years) were recruited into the study. The 152 healthy controls (HC) (68 males and 84 females, mean age 31.01 ± 9.51 years), matched for age, sex, and education level, were enrolled from the local area through internet advertisements, and screened by clinicians using the non-patient version of SCID to confirm absence of Axis I psychiatric disorders. The HC were excluded if they had a first-degree relative with a history of psychiatric disorders. This study was approved by the Research Ethics Committee of West China Hospital of Sichuan University, and the written fully-informed consent was obtained from all participants. MRI data acquisition MRI scanning was performed on a 3 T MRI scanner (Trio Tim, Siemens AG, Erlangen, Germany) with an 8-channel phased array head coil. Participants were positioned comfortably in the coil, fitted with earplugs and foam pads, and instructed to keep their eyes closed, keep still, and not think of anything specific. Resting-state functional MRI (rs-fMRI) data were acquired using a gradient-echo echo-planar imaging sequence with repetition time (TR)/ echo time (TE) = 2000/30 ms, flip angle = 90°, 30 axial slices, 5 mm slice thickness with no slice gap, field of view (FOV) = 240 × 240 mm2, matrix = 64 × 64, and voxel size = 3.75 × 3.75 × 5 mm2. The total acquisition time was around 350 seconds, containing 175 volumes. High-resolution T1-weighted anatomical images were obtained in sagittal orientation using a three-dimensional magnetization prepared rapid acquisition gradient echo sequence with TR/ TE = 1900/2.26 ms, inversion time (TI) = 900 ms, flip angle = 9°, 176 contiguous sagittal slices, 1 mm slice thickness, FOV = 256 × 256 mm2, matrix =256 × 256, and voxel size = 1 × 1 × 1 mm3. MRI data preprocessing The rs-fMRI data was preprocessed through a standardized pipeline in the Data Processing and Analysis for Brain Imaging toolkit (DPABI, version 7.0, http://www.restfmri.net) [34]. The first five volumes were discarded to ensure signal stabilization, and a slice timing correction applied, then the time series of images were realigned using a six-parameter (rigid body) linear transformation. Realigned T1-weighted images were co-registered to the mean functional image and segmented into gray matter, white matter and cerebrospinal fluid. The Friston 24-parameter model was used to regress out head motion confounding effects, including 6 head motion parameters, previous time point of 6 head motion parameters, and 12 corresponding squared items [35]. The mean framewise displacement (FD) was calculated for each subject as a measure of microscale head motion: data was included only for participants with mean FD < 0.25 mm, and mean FD is regressed as a covariate in subsequent statistical analyses [36]. Next, several nuisance signals (white matter and cerebrospinal fluid signals) were regressed out to reduce the influences of non-neuronal blood oxygen level-dependent fluctuations. Then the images were spatially normalized to standard Montreal Neurological Institute (MNI) template (voxel size 3 × 3 × 3 mm3) and spatially smoothed with a 6 mm full width at half maximum (FWHM) Gaussian kernel. Finally, temporal bandpass filtering (0.01-0.08 Hz) was performed to account for low-frequency machine magnetic field drifts and high-frequency respiratory or cardiac noise, and the effects of systematic drifts or trends in fMRI were removed by detrending analysis. Connectivity-based parcellation The CBPtools package [31], implemented in Python 3.7, was used to segment the ACC into distinct functional subregions per hemisphere, accounting for potential lateralization effects [1]. The CBP method is described in detail elsewhere [30, 31, 37] and summarized as a flowchart in Fig. 1. Briefly, resting-state functional connectivity (rsFC) was computed between each voxel within the left and right ACC region (seed mask, from the AAL2) and every voxel in the rest of whole-brain gray matter (targeted mask, the MNI 152 T1 whole-brain gray-matter template) to obtain a voxel-wise seed-to-target connectivity matrix for each subject. The target mask was subsampled to improve computational efficiency, and seed voxels were removed from the target mask to avoid domination of clustering by typically high intra-ACC. The group-level consensus parcellation was generated across all subjects to underpin between-group comparisons. To obtain individual-level parcellation of the ACC, k-means clustering was applied to each subject’s connectivity matrix to assign each voxel within the ACC to a cluster, grouping similar voxels based on their connectivity patterns. Main parameters of the k-means algorithm were set as follows: k from 2 to 5, k-means + + initialization, 256 initializations, and a maximum of 10,000 iterations. Our choice of k-means clustering was informed by previous studies, in which it has been widely adopted for its efficiency, scalability, and suitability for high-dimensional neuroimaging data [29, 30]. The ranging of k values from 2 to 5 was guided by anatomical and functional parcellation research on the ACC. For instance, Margulies et al. delineated five ACC subregional coordinates (subgenual, pregenual, rostral, dorsal, and caudal) that commonly used in MDD-related studies [25, 26]. Other studies proposed two (ventral and dorsal) or three (subgenual, pregenual, and supragenual) ACC partitions in healthy individuals [27,15]. Accordingly, we chose k = 2-5 to ensure compatibility with established ACC parcellation models. Since the arbitrary cluster labels were assigned at individual level, hierarchical clustering with complete linkage and Hamming distance was performed on individual clustering for each k to relabel clusters consistently across subjects, and the Hamming distance was applied to account for inherent arbitrariness of the cluster labels [31]. Group-level clustering was achieved by calculating the mode of individual clustering labels for each voxel within the ACC that best describes all included subjects for each k. The optimal number of clusters was assessed using three cluster quality metrics: the Silhouette index, the Davies-Bouldin index, and the Calinski-Harabasz index. We further computed the adjusted rand index as a measure of similarity, and the relabeling accuracy between the individual- and the group-level clustering.Fig. 1Flowchart of connectivity-based parcellation (CBP) procedure.ACC anterior cingulate cortex, AAL automated anatomical labeling atlas, MNI Montreal Neurological Institute, GM gray matter, FWHM full width at half maximum. Flowchart of connectivity-based parcellation (CBP) procedure. ACC anterior cingulate cortex, AAL automated anatomical labeling atlas, MNI Montreal Neurological Institute, GM gray matter, FWHM full width at half maximum. Functional connectivity analysis The CBP-derived ACC subregions from each hemisphere were used as seeds for functional connectivity analysis using the Resting-State fMRI Data Analysis Toolkit (RESTplus version 1.24, http://restingfmri.sourceforge.net). To acquire the subject-level rsFC maps, regional time series within each seed (averaging across all voxels) were extracted from the preprocessed data and then correlated (using Pearson’s correlation) with the time series of every voxel in the whole brain. Correlation coefficients were standardized using a Fisher r-to-z transformation to obtain the z-value rsFC maps per subject, to improve the normality for further statistical analyses. Statistical analysis Group-level statistical analyses were conducted in SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/). Voxel-wise diagnosis-by-subregion flexible-factorial analysis of variance (ANOVA) was performed in left and right ACC seeds separately, with diagnosis as a between-group factor and subregion as a within-subject factor, and with age, sex, education, and head motion as covariates to account for their possible confounding effects. To determine group difference in differentiated ACC subregional connectivity patterns, we calculated the diagnosis × subregion interaction effects. We also examined main effects of diagnosis on ACC connectivity. Furthermore, we explored distinctive rsFC profiles between ACC subdivisions per hemisphere within all participants, in MDD patients and in HC subjects, respectively. Statistical significance threshold was set as P < 0.005 at the voxel level and family-wise-error (FWE) corrected P < 0.025 (0.05/2, subregional ACC seeds from two hemispheres) at the cluster level, both two-sided. We labeled resulting region with the AAL3 as a reference. The rsFC values were extracted from the clusters showing significant diagnosis-by-subregion interactions for post hoc analysis in SPSS 25.0 using the simple effects test, with Bonferroni-corrected two-sided P < 0.05 as the statistical significance. After linear correlation conditional tests (e.g., Kolmogorov-Smirnov normal distribution test), Spearman correlation was used to explore relationships between the rsFC alterations of ACC subregions with clinical features (namely each-item and total score of HAMD-17, total score of HAMA-14, illness duration, and age of onset) in MDD patients. These heuristic correlation analyses were presented using nominal significance thresholds (two-sided P < 0.05). Validation and reproducibility To evaluate the robustness of the CBP-derived ACC parcellations, we replicated the CBP processing for ACC separately in MDD and HC populations, and used adjusted rand index and Dice coefficient to quantify the consistency of ACC parcellations between the full sample, the MDD, and the HC groups. We further validated results using two alternative clustering methods, hierarchical and spectral clustering, available in the CBPtools package. A systematic cross-method comparison for k = 2-5 cluster solutions was performed across multiple metrics of internal validity (Silhouette, Calinski-Harabasz, and Davies-Bouldin indices) and clustering quality (adjusted rand index and relabeling accuracy). We also tested the effects of using several alternative preprocessing steps: preprocessing with 4 mm and 8 mm FWHM Gaussian kernel smoothing, performing global signal regression (GSR), and using a more stringent head motion control strategy, namely motion scrubbing with a FD threshold of 0.50 mm, excluding subjects with < 5 min of scrubbed data: no participant failed this criterion, on average only 6 frames being scrubbed per subject (3.69%). Results Sociodemographic and clinical characteristics The sociodemographic and clinical characteristics of study participants were shown in Table 1. There was no significant difference in age, sex, education and head motion between two groups (P > 0.05). Mean illness duration was around 6 months and age of onset was about 30 years in MDD. Total HAMD-17 score was 24.72 ± 5.38, corresponding to moderate and severe depressed symptoms, and total HAMA-14 score was 23.35 ± 8.54.Table 1Sociodemographic and clinical characteristics of the study participants.MDD (N = 158)HC (N = 152)P valueAge, years29.77 ± 9.8231.01 ± 9.510.259Sex (female/male)99/5984/680.186Education, years--0.797Middle school and lower2732-High school2614-University and higher105106-Mean FD, mm0.112 ± 0.0490.122 ± 0.0520.093BMI21.06 ± 4.00--Illness duration, weeks26.40 ± 26.80--Age of onset, years29.26 ± 9.84--HAMD-1724.72 ± 5.38--HAMA-1423.35 ± 8.54--Unless otherwise indicated, data are presented as mean ± standard deviation.MDD major depressive disorder, HC healthy controls, BMI body mass index, FD framewise displacement, HAMD-17 17-item hamilton depression rating scale, HAMA-14 14-item hamilton anxiety rating scale. Table 1Sociodemographic and clinical characteristics of the study participants.MDD (N = 158)HC (N = 152)P valueAge, years29.77 ± 9.8231.01 ± 9.510.259Sex (female/male)99/5984/680.186Education, years--0.797Middle school and lower2732-High school2614-University and higher105106-Mean FD, mm0.112 ± 0.0490.122 ± 0.0520.093BMI21.06 ± 4.00--Illness duration, weeks26.40 ± 26.80--Age of onset, years29.26 ± 9.84--HAMD-1724.72 ± 5.38--HAMA-1423.35 ± 8.54--Unless otherwise indicated, data are presented as mean ± standard deviation.MDD major depressive disorder, HC healthy controls, BMI body mass index, FD framewise displacement, HAMD-17 17-item hamilton depression rating scale, HAMA-14 14-item hamilton anxiety rating scale. Sociodemographic and clinical characteristics of the study participants. Unless otherwise indicated, data are presented as mean ± standard deviation. MDD major depressive disorder, HC healthy controls, BMI body mass index, FD framewise displacement, HAMD-17 17-item hamilton depression rating scale, HAMA-14 14-item hamilton anxiety rating scale. Summary of main findings The data-driven CBP approach identified two ACC functional subregions: the vACC, shown exclusive connections to DMN regions, and the dACC, primarily connected with salience network and subcortical areas (hypothesis 1). Diagnosis-by-subregion interactions revealed an imbalanced ventral/dorsal ACC-DMN connectivity pattern specific to MDD (hypothesis 2). Moreover, vACC-DMN hypoconnectivity was related with lower mood and longer illness duration, while dACC-DMN hyperconnectivity was linked to greater suicidality and insomnia (hypothesis 3). Identification of ACC functional subregions using CBP Using CBP we identified two functional subregions of the ACC in each hemisphere, namely ventral ACC (vACC) and dorsal ACC (dACC), as the optimal parcellation among 2- to 5-cluster solutions (Fig. 2A), according to three internal validity metrics (Fig. S1). This two-cluster parcellation exhibited acceptable similarity and relabeling accuracy between the individual-level and the group-level clustering results (Fig. S2). We illustrated the distinctive rsFC patterns between ventral and dorsal ACC in all participants taken together (Fig. 2B), in the MDD patients alone (Fig. 2C), and in the HC subjects alone (Fig. 2D); connectivity patterns in the three groups were rather similar. The two ACC subregions showed different intrinsic connectivity profiles: the vACC had stronger connectivity than the dACC (‘hot’ colors in Fig. 2B-D) with the regions of default mode network (DMN), namely dorsomedial prefrontal cortex (dmPFC), posterior cingulate cortex (PCC), inferior parietal and temporal lobes; by contrast the dACC had stronger connectivity than the vACC (‘cool’ colors in Fig. 2B-D) with the regions of salience network and subcortex, namely anterior insula, middle cingulate cortex extending to supplementary motor area, temporoparietal junction, lateral PFC, striatum and thalamus.Fig. 2Connectivity-based parcellation of the ACC and its distinctive subregional functional connectivity patterns.A shows the two-cluster solution derived by the CBP in the whole subject group: dACC and vACC (detailed in the insets) are colour-coded as in the key; left and right shown separately. The next three panels compare the connectivity of the dorsal and ventral ACC, shown in left and right view and axial section: vACC showed higher connectivity with the default mode network, whereas dACC showed higher connectivity with the salience network and subcortical regions in the whole subject group (B), in HC alone (C), and in MDD patients alone (D): warm colors (red) indicate that vACC connectivity exceeds dACC connectivity, while cool colors (blue) indicates that dACC connectivity exceeds vACC connectivity. CBP connectivity-based parcellation, ACC anterior cingulate cortex, vACC ventral ACC, dACC dorsal ACC, rsFC resting-state functional connectivity, MDD major depressive disorder, HC healthy control. Connectivity-based parcellation of the ACC and its distinctive subregional functional connectivity patterns. A shows the two-cluster solution derived by the CBP in the whole subject group: dACC and vACC (detailed in the insets) are colour-coded as in the key; left and right shown separately. The next three panels compare the connectivity of the dorsal and ventral ACC, shown in left and right view and axial section: vACC showed higher connectivity with the default mode network, whereas dACC showed higher connectivity with the salience network and subcortical regions in the whole subject group (B), in HC alone (C), and in MDD patients alone (D): warm colors (red) indicate that vACC connectivity exceeds dACC connectivity, while cool colors (blue) indicates that dACC connectivity exceeds vACC connectivity. CBP connectivity-based parcellation, ACC anterior cingulate cortex, vACC ventral ACC, dACC dorsal ACC, rsFC resting-state functional connectivity, MDD major depressive disorder, HC healthy control. Subregion-specific ACC-DMN dysconnectivity in MDD Voxel-wise flexible-factorial ANOVA revealed significant diagnosis (MDD vs HC) × subregion (vACC vs dACC) interactions on ACC connectivity, primarily in several regions of the DMN (Fig. 3A). Specifically, both left and right ACC connectivity showed significant interactions in bilateral PCC (left ACC: F = 28.92, PFWE-corrected < 0.001; right ACC: F = 29.97, PFWE-corrected < 0.001) and left middle temporal gyrus (MTG; left ACC: F = 16.61, PFWE-corrected = 0.0180; right ACC: F = 19.96, PFWE-corrected = 0.0246). Right ACC connectivity showed significant interactions in additional regions, including bilateral dmPFC (F = 12.99, PFWE-corrected = 0.0124), left inferior parietal lobule (IPL, F = 16.95, PFWE-corrected = 0.0246) and right superior parietal lobule (SPL, F = 20.39, PFWE-corrected = 0.0034) (Fig. 3A and Table S1). Post hoc analyses further revealed distinct subregional connectivity patterns: for the left ACC, the vACC exhibited significantly decreased connectivity with bilateral PCC and left MTG in MDD vs HC, whereas the dACC showed significantly increased connectivity with bilateral PCC and moderately increased connectivity with left MTG. Similarly, for the right ACC, the vACC showed significantly reduced connectivity with bilateral PCC, left MTG, bilateral dmPFC, and left IPL, in MDD vs HC, while the dACC connectivity with these regions was modestly elevated. In addition, we found a relative enhancement of connectivity (i.e., a decrease in negative connectivity) between right vACC and right SPL in MDD vs HC, alongside a slight reduction in connectivity between right dACC and right SPL. These results including statistical details (e.g., P-values, effect sizes and confidence intervals) were illustrated in Fig. 3B, C and detailed in Tables S2-S3. We also explored main effect of diagnosis on the ACC connectivity. MDD patients showed enhanced connectivity between bilateral ACC and right orbitofrontal cortex (left ACC: t = 4.83, PFWE-corrected < 0.001; right ACC: t = 4.99, PFWE-corrected = 0.0014), and between right ACC with both right ventrolateral PFC (t = 4.19, PFWE-corrected = 0.0069) and right opercular part of inferior frontal gyrus (t = 4.06, PFWE-corrected < 0.001), compared to HC (Fig. S4).Fig. 3Altered subregional ACC connectivity in MDD compared with HC.A shows brain regions (labelled below the images) with significant interaction between diagnosis (MDD vs HC) and subregion (vACC vs dACC) in ACC subregional connectivity; left and right are shown separately. Panel (B) shows the significant interactions between diagnosis and subregion computed separately in left and right ACC connectivity as line-graphs (y-axis is rsFC, x-axis groups are MDD and HC; error bars represent standard errors of means; each sub-graph aligns with the corresponding brain regions above). C shows the results of post hoc simple effects tests comparing ACC connectivity between MDD versus HC in each subregion as labelled (each graph aligns with the corresponding brain regions above) as bar-charts (y-axis is rsFC, x-axis groups are vACC and dACC; error bars indicate standard deviation; each sub-graph aligns with the corresponding brain regions above). *P < 0.05, **P < 0.01, ***P < 0.005, Bonferroni corrected. MDD major depressive disorder, HC healthy control, rsFC resting-state functional connectivity, ACC anterior cingulate cortex, vACC ventral ACC, dACC dorsal ACC, PCC posterior cingulate cortex, MTG middle temporal gyrus, dmPFC dorsomedial prefrontal cortex, IPL inferior parietal lobule, SPL superior parietal lobule. Altered subregional ACC connectivity in MDD compared with HC. A shows brain regions (labelled below the images) with significant interaction between diagnosis (MDD vs HC) and subregion (vACC vs dACC) in ACC subregional connectivity; left and right are shown separately. Panel (B) shows the significant interactions between diagnosis and subregion computed separately in left and right ACC connectivity as line-graphs (y-axis is rsFC, x-axis groups are MDD and HC; error bars represent standard errors of means; each sub-graph aligns with the corresponding brain regions above). C shows the results of post hoc simple effects tests comparing ACC connectivity between MDD versus HC in each subregion as labelled (each graph aligns with the corresponding brain regions above) as bar-charts (y-axis is rsFC, x-axis groups are vACC and dACC; error bars indicate standard deviation; each sub-graph aligns with the corresponding brain regions above). *P < 0.05, **P < 0.01, ***P < 0.005, Bonferroni corrected. MDD major depressive disorder, HC healthy control, rsFC resting-state functional connectivity, ACC anterior cingulate cortex, vACC ventral ACC, dACC dorsal ACC, PCC posterior cingulate cortex, MTG middle temporal gyrus, dmPFC dorsomedial prefrontal cortex, IPL inferior parietal lobule, SPL superior parietal lobule. Clinical correlations of ACC subregional dysconnectivity The lower connectivity of right vACC and bilateral PCC was correlated to longer illness duration (r = -0.194, P = 0.015), and both the higher connectivity of right vACC and right SPL (r = 0.166, P = 0.037) and the lower connectivity of right vACC and left IPL (r = -0.168, P = 0.035) were associated with more severe depressed mood (HAMD-1) (Fig. 4A). Moreover, suicide (HAMD-3) was positively correlated to the connectivity between left dACC and bilateral PCC (r = 0.164, P = 0.040) and between bilateral dACC and left MTG (left ACC: r = 0.186, P = 0.022; right ACC: r = 0.169, P = 0.033), and insomnia (HAMD-5) was positively associated with the connectivity between right dACC with left IPL (r = 0.167, P = 0.036) and bilateral dmPFC (r = 0.165, P = 0.038) (Fig. 4B).Fig. 4Correlations between clinical symptoms and ACC connectivity, showing significant diagnosis-by-subregion interactions.A shows correlations of vACC dysconnectivity with core depressive symptoms, and (B) shows correlations of dACC dysconnectivity with maladaptive behaviors. Upper right schematics demonstrate subregional ACC dysconnectivity with relevance to clinical variables. The seeds of ACC subregions are shown in the black box. ACC anterior cingulate cortex, vACC ventral ACC, dACC dorsal ACC, rsFC, resting-state functional connectivity, PCC posterior cingulate cortex, MTG middle temporal gyrus, SPL superior parietal lobule, IPL inferior parietal lobule, dmPFC dorsomedial prefrontal cortex, HAMD Hamilton Depression Rating Scale. Correlations between clinical symptoms and ACC connectivity, showing significant diagnosis-by-subregion interactions. A shows correlations of vACC dysconnectivity with core depressive symptoms, and (B) shows correlations of dACC dysconnectivity with maladaptive behaviors. Upper right schematics demonstrate subregional ACC dysconnectivity with relevance to clinical variables. The seeds of ACC subregions are shown in the black box. ACC anterior cingulate cortex, vACC ventral ACC, dACC dorsal ACC, rsFC, resting-state functional connectivity, PCC posterior cingulate cortex, MTG middle temporal gyrus, SPL superior parietal lobule, IPL inferior parietal lobule, dmPFC dorsomedial prefrontal cortex, HAMD Hamilton Depression Rating Scale. Validation and reproducibility The two dorsoventral ACC functional subregions derived from the whole participants showed highly similar spatial extents and subregional connectivity patterns to those derived within the MDD or the HC group separately (Fig. S3). All the adjusted rand index and Dice coefficient values exceeded 0.80 across groups, and spatial similarity of distinct connectivity patterns based on ACC subregions obtained from different groups exceeded 0.85 (Table. S4). The CBP-derived ACC parcellations from alternative clustering methods of hierarchical (Fig. S5) and spectral clustering (Fig. S6) accorded well with the main findings. As shown in Supplementary Fig. S7, k-means clustering consistently outperformed the other methods across nearly all internal validity and clustering quality metrics. The results mostly remained consistent whether the fMRI data was smoothed with the 4 mm or the 8 mm FWHM Gaussian kernel (Fig. S8 and S9), and were largely unchanged using preprocessed images with global signal regression (Fig. S10) or using the aggressive head motion control strategy of scrubbing (Fig. S11). Discussion Using the data-driven CBP approach, we established that, in the context of MDD, the ACC was best represented by a bipartite ventral-dorsal functional subdivision in each hemisphere, as judged by three clustering quality metrics. These two subregions had distinct intrinsic connectivity profiles: the vACC was preferentially connected with DMN regions, while the dACC primarily interacted with salience network and subcortical areas. Based on these CBP-defined ACC functional subregions, we revealed for the first time an imbalanced ventral/dorsal ACC-DMN connectivity pattern specific to MDD through diagnosis-by-subregion interaction analyses. In HC, the vACC demonstrated stronger connectivity with key components of the DMN (i.e., PCC, MTG, dmPFC and IPL) compared to the dACC. However, this difference between vACC- and dACC-DMN connectivity was attenuated in MDD, driven by the vACC hypoconnectivity and a trend towards dACC hyperconnectivity with these DMN regions. Furthermore, lower vACC-DMN connectivity was linked to lower mood and longer illness duration, whereas higher dACC-DMN connectivity was related with greater suicidal risk and insomnia. These findings underscored the ventral-dorsal functional heterogeneity of ACC architecture and revealed the symptom-specific alterations in subregional ACC-DMN connectivity in MDD, which may potentially serve as biomarkers for divergent symptom profiles and help guide tailored interventions. The first stage of this work, which underpinned the findings on altered ACC subregional connectivity in MDD, was functional parcellation for the ACC based on CBP. This connectivity-based framework with data-driven clustering algorithms yielded an optimal bipartite dorsoventral ACC functional partition that best fit to our data (Fig. 2A). This broadly aligns with anatomically-defined bipartite ACC segmentations [23], which presumably reflect neural network interactions depending on anatomical neuron properties such as size, arrangement and configuration [38]. While some studies proposed tripartite or more complex ACC subdivisions, these have mainly been derived from geometric, anatomical, or tractography-based approaches [26, 27, 39], which may not reflect true functional boundaries, and lead to substantial overlap in intrinsic connectivity profiles, especially between adjacent subregions. Unlike traditional structural partitions [28, 29], the CBP enables a fully data-driven, voxel-wise functional parcellation according to intrinsic connectivity properties, allowing better population-specific characterization of ACC functional architecture and improved detection of subtle subregional connectivity changes [30, 31]. A prior study in healthy subjects identified comparable bipartite ACC partitions (termed pregenual and supracallosal) using voxel-to-region clustering based on connectivity with AAL2-based regions [15]. Extending this work, our study incorporated clinical populations and applied a voxel-to-voxel CBP framework to delineate functionally-informed ACC subregions and their distinct connectivity patterns in the context of MDD. Moreover, we validated the robustness of the CBP-defined ACC functional partitions using hierarchical and spectral clustering, both yielding highly consistent results. The two ACC subdivisions showed preferential intrinsic connectivity pattern with specialized target areas: the vACC showed exclusive connections with DMN regions, while the dACC was predominantly connected to salience network and subcortical areas (Fig. 2B-D). These patterns remained robust across alternate preprocessing methods, including different kernel smoothing, global signal regression, and aggressive motion scrubbing (Fig. S8-S11). This accorded with a report in HC of stronger ventral ACC connectivity with medial PFC and PCC (DMN), while stronger dorsal ACC connectivity with insula (salience network), putamen and pallidum (subcortical areas) [15]. These findings corroborated earlier research suggesting that the vACC was involved in affective processing through connections with emotion-related areas (e.g., DMN and limbic circuits), while the dACC was engaged in cognitive functions through connections to regions related to attention and cognition (e.g., frontoparietal and subcortical regions) [25, 26]. This confirmed the well-established gradual transition in the functional roles of ventral-dorsal ACC partitions, shifting from “affective” to “cognitive” processes [23, 24]. Comparing to prior research, based on this CBP-defined bipartite ACC functional partition, we revealed for the first time an imbalanced pattern of ventral/dorsal ACC-DMN connectivity linked to distinct clinical manifestations in MDD. Specifically, the vACC-DMN hypoconnectivity was related to lower mood and longer illness duration, while the dACC-DMN hyperconnectivity was related to higher level of suicide and insomnia. This functional dissociation has not been reported in studies using anatomical or literature-based ACC definitions, perhaps because those methods averaged across functionally heterogeneous regions, obscuring functional subregion-specific effects. In HC, the vACC connectivity with DMN regions (namely PCC, dmPFC, MTG and IPL) was stronger than that of dACC, which may facilitate efficient processing of self-referential emotional information and social cognition crucial for maintaining emotional health and balanced mental states [40, 41]. In contrast, MDD showed marked attenuation of this difference, driven by the hypoconnectivity of vACC combined with a trend towards hyperconnectivity of dACC with these DMN regions. The vACC, an affective ACC division and prominent anterior midline core of DMN, mediates self-relevant emotional modulation and affective decision [24, 42]. Reduced connectivity between vACC and other DMN regions may reflect disrupted communications within the DMN and dysfunction in maintaining normal emotional response that further leads to core depressive symptoms, such as negative self-referential emotion and impaired self-representation [43–45]. The dorsal ACC, a cognitive ACC division and key locus of salience network, integrates interoceptive signals and external stimuli, facilitates switching of DMN and executive network [46], playing a major role in cognitive control of negative bias, action appraisal, and adaptive goal-directed behaviors [24, 47, 48]. Elevated connectivity between dACC and DMN had been linked to greater susceptibility to self-related negative emotion and inability to disengage from pessimistic thoughts that further lead to maladaptive behaviors, such as suicide and insomnia [49, 50]. Specifically, MDD patients showed altered ACC subregional connectivity with bilateral PCC (i.e., vACC-DMN hypoconnectivity and dACC-DMN hyperconnectivity). Both PCC and vACC were midline cores of DMN critically involved in self-referential affective processing [42]. Disrupted vACC-PCC connectivity corroborated meta-analytic evidence [16, 43], and aligned with decreased structural connectivity between anterior and posterior parts of DMN cores [51]. This disruption may underlie impaired ability to modulate self-relevant negative emotion, persistently high levels of depressive affection, and trait rumination [52, 53]. Conversely, enhanced dACC-DMN connectivity likely reflected excessive focus on internal negative thoughts and difficulty shifting attention toward external goals. This aligned with findings of heightened dACC and PCC activations in response to negative distractors, and a positive correlation between dACC-PCC connectivity and depressive severity during affective interference, suggesting disrupted coordination between cognitive control and self-focused attention in MDD [54]. MDD patients showed reduced connectivity between bilateral vACC and left MTG, a region within the DMPFC subsystem of DMN involved in mentalization and meta-cognition [42, 55]. This vACC-MTG (midline core and DMPFC subsystem) dysconnectivity likely contributed to deficits in perception of own or others’ mental states and social cognition in MDD. Our findings aligned with ENIGMA reports of cortical thinning in cingulate cortex, insula and lateral temporal lobes, more pronounced in first-episode adult-onset patients [11], and with a recent meta-analytic finding of reduced vACC-MTG connectivity in MDD [16]. We also found decreased connectivity between right vACC and bilateral dmPFC (another part of the DMPFC subsystem), indicating dysfunction in self-related decision-making, present-moment awareness, and emotional regulation in MDD [4, 42]. This accords with reports of MDD-related structural and functional deficits in ACC subregions and medial/lateral PFC [8], such as smaller volume or thinner thickness [11, 13] and loss of global connectivity in medial PFC extending to vACC [56–59], which may reflect self-representation deficit and emotional dysregulation in MDD. Moreover, MDD patients exhibited lower connectivity of right vACC and left IPL and relatively higher connectivity of right vACC and right SPL. Disrupted connectivity of vACC (anterior DMN core) and left IPL (posterior MTL subsystem) in MDD may underlie negative conscious sense-of-self and passive expectation for the future [53, 60]. Given the role of right SPL (key locus of frontoparietal and attentional networks) in decision-making and attention reorientation [61, 62], the relatively higher vACC-SPL connectivity may suggest an excessive internal focus rather than external attention in MDD, in accordance with enhanced interactions between DMN and frontoparietal areas [63]. These subregion-specific ACC dysconnectivity patterns had relevance to divergent clinical manifestations of MDD, which had not been detected in studies using conventional ACC partitions. The vACC-centered hypoconnectivity patterns were associated with core depressive symptoms. Lower vACC-PCC connectivity was correlated with prolonged illness duration, suggesting an increasing connectivity dissociation between vACC and PCC as the depressive state persisted [64]. Additionally, lower vACC-IPL connectivity and higher vACC-SPL connectivity were correlated with more severe depressed mood, potentially reflecting impaired self-perception and excessively negative self-concerns. Conversely, the dACC-centered hyperconnectivity patterns were linked to elevated maladaptive behaviors, with suicide positively related to dACC-PCC/MTG connectivity, and insomnia positively related to dACC-dmPFC/IPL connectivity. An earlier study reported morphological changes in the ACC and PCC related to increased suicidal risk [65], and another observed higher dACC-PCC coupling in suicidal ideators than controls [66]. Large-sample studies have uncovered structural deficits in frontal and parietal regions in MDD patients with severe insomnia [67], alongside heightened connectomes in prefrontal, sensorimotor and parietal cortices mediating the relationship between depressive problems and poor sleep quality [68]. Several limitations merit consideration and guide future directions. First, while excluding patients on medication and with comorbidity facilitated to identify MDD-specific abnormalities, it reduced the generalizability of our findings. Future studies involving more diverse MDD populations are required to validate and extend our results. Second, correlations between connectivity changes and clinical features that did not survive correction for multiple comparisons should be considered preliminary. They were, however, suggestive of subtle but important links between subregional ACC-DMN dysconnectivity and specific symptom profiles in MDD that merit validation in larger samples using hypothesis-driven, multivariate models. Finally, although using group-specific CBP in our analysis may limit the universality of our results, it surpasses the standard structural templates in providing better representation for group-level functional connectivity, particularly given the cross-study inconsistencies in subject characteristics, scanning protocols, and preprocessing strategies. Future research should prioritize the individualized brain segmentation, e.g. training machine learning models on large and high-quality independent datasets to predict or define the ACC partitions at the individual level. In conclusion, utilizing a data-driven CBP technique, we parcellated the ACC into a bipartite ventral-dorsal functional subdivision in each hemisphere (namely vACC and dACC) with distinct subregional intrinsic connectivity profiles. The vACC showed exclusive connections with DMN regions, while the dACC was predominantly connected to salience network and subcortical areas. In HC, the vACC exhibited stronger connectivity with key components of DMN compared to the dACC, while in MDD this difference was attenuated, driven by the vACC hypoconnectivity and a trend towards dACC hyperconnectivity with these DMN regions. Clinically, lower vACC-DMN connectivity was related to core depressive symptoms, i.e., lower mood and longer illness duration, whereas higher dACC-DMN connectivity was linked to more maladaptive behaviors, e.g., suicide and insomnia. Our findings confirmed ventral-dorsal functional heterogeneity of ACC architecture and elucidated its symptom-specific subregional dysconnectivity patterns with DMN regions in MDD. This may provide insights into developing tailored interventions for divergent clinical manifestations. For instance, modulating the vACC may be more effective in alleviating core depressive symptoms, whereas targeting the dACC may preferentially ameliorate maladaptive behaviors. Future research should validate these findings in longitudinal designs and explore their utility in personalized therapeutic strategies. Supplementary information Supplementary Material Supplementary Material