Heterogeneous brain alterations in treatment-resistant depression converge on common control networks

Abstract

Treatment-resistant depression (TRD) imposes a substantial clinical and societal burden. However, it remains unclear whether neuroimaging abnormalities associated with TRD converge onto specific large-scale brain networks. We systematically reviewed neuroimaging articles comparing patients with TRD to both healthy controls (HCs) and patients with non-treatment-resistant depression (non-TRD), encompassing functional and structural imaging modalities. Using a validated coordinate network mapping (CNM) approach and a normative human brain connectome, we derived TRD-associated networks. Overlap analyses with canonical large-scale networks were performed to delineate the functional architecture of these networks. In addition, spatial correlations between TRD-associated probability maps and whole-brain neurotransmitters density maps to explore potential neurochemical substrates. In the TRD–HC contrast, TRD-associated functional and structural networks were distributed across frontal, cingulate, insular, striatal, and thalamic regions, showing their greatest overlap with the frontoparietal (FPN) and cingulo-opercular network (CON). In the TRD–non-TRD contrast, TRD-associated functional and structural networks were more selectively localized, with predominant overlap confined to the CON. Spatial correlation analyses further anchored these network-level abnormalities to glutamatergic, dopaminergic, and monoaminergic neurotransmitter systems. These findings indicate that the general neurobiological burden of TRD is anchored within large-scale control networks involving the FPN and CON, whereas treatment resistance is characterized by more focal alterations within the CON. This network-based dissociation provides insight into the neural organization of TRD and highlights potential circuit-level targets for therapeutic intervention.

Introduction

Major depressive disorder (MDD) is a recurrent and often chronic psychiatric illness. Despite the availability of multiple antidepressant treatments, a substantial proportion of patients fail to achieve adequate clinical improvement, thereby meeting criteria for treatment-resistant depression (TRD). Approximately 20–30% of patients with MDD develop TRD, typically defined as the failure to achieve clinical response following at least two adequate antidepressant trials [1,2]. TRD is associated with markedly worse prognosis, elevated suicide risk, and a disproportionate healthcare and functional burden [3,4]. Some evidence suggests that TRD may constitute a distinct subtype of MDD, rather than simply reflecting greater illness severity [5,6]. Whether TRD represents a discrete subtype or an advanced stage, delineating the neurobiological features that underlie treatment resistance is critical, as such insights may facilitate earlier identification and inform the development of more effective interventions.

Neuroimaging research increasingly indicates that depression is characterized by disruptions in large-scale brain networks, rather than isolated regional abnormalities [7]. Across recurrent episodes, structural and functional alterations have been observed in circuits centered on the dorsolateral prefrontal cortex (dlPFC), anterior cingulate cortex (ACC), insula, hippocampus, and amygdala, suggesting a cumulative neural burden over the course of illness [[8], [9], [10]]. Importantly, abnormalities within several of these regions, particularly those embedded within the default mode network (DMN), have emerged as potential signatures distinguishing treatment-responsive from treatment-resistant depression [5,11,12]. Beyond these regional observations, TRD is further characterized by widespread network-level disruptions, including altered interactions among sensory, executive, and affective regulatory systems [5,12]. Consistent with this network perspective, neuromodulatory interventions such as repetitive transcranial magnetic stimulation (TMS) and deep brain stimulation (DBS) have been developed to directly target circuits implicated in TRD, with demonstrated therapeutic benefit in patients [[13], [14], [15]]. Probabilistic tractography has further shown that anatomically disparate DBS targets converge on overlapping white-matter pathways, with Brodmann area 10 and the amygdala emerging as key hubs linking these targets [16].

Despite these advances, progress toward understanding the neurobiology of TRD and the mechanisms through which neuromodulatory interventions exert their therapeutic effects has remained limited, in part because neuroimaging studies of TRD have yielded highly heterogeneous findings [17]. This lack of convergence has traditionally been interpreted as indicating that only brain regions repeatedly identified across studies are reliable or disease-relevant, with variability attributed to clinical or methodological differences. However, recent work by Michael D. Fox and colleagues challenges this assumption, proposing that seemingly inconsistent regional findings may connect to a shared network [18]. Under this framework, the implicated network may be spatially distant from the regions reported in individual studies, yet functionally connected to them, suggesting that regional heterogeneity may converge at the level of network architecture [19]. Supporting this view, a recent coordinate network mapping (CNM) study demonstrated that structurally heterogeneous findings in MDD converge onto the frontoparietal (FPN), dorsal attention (DAN), and visual networks (VN), highlighting shared depression-related circuits across younger and older adults [20].

Few systematic reviews and only one meta-analysis have summarized neuroimaging findings in TRD, underscoring that its neural characteristics remain comparatively underexplored [5,7,21]. The present investigation aims to address two questions regarding TRD-related brain alterations: first, which neural abnormalities reflect the broader neurobiological burden of TRD? And second, what specific features mark resistance to conventional pharmacotherapy? Clarifying these issues is essential for defining the neurobiological signature of TRD and for guiding earlier detection and more effective treatment strategies.

To address these questions, we directly compared patients with TRD to both healthy controls (HCs) and patients who are not treatment resistant (non-TRD) in terms of functional and structural brain characteristics. We conducted CNM analysis using the Human Connectome Project dataset (mean age = 21.40 ± 2.86 years; 500 males and 500 females) to identify convergent network-level patterns underlying heterogeneous neuroimaging findings in TRD. To further characterize the composition of these networks, we examined their overlap with canonical functional systems using the 12 cortical–subcortical networks defined by Ji et al. [22]. In addition, we assessed cross-modality correspondence to assess the consistency of structural and functional alterations, and cross-contrast correspondence to dissociate abnormalities reflecting general disease burden from those specifically associated with treatment resistance. Finally, spatial correlations with neurotransmitter maps were examined to probe potential underlying neurochemical mechanisms (workflow shown in Fig. 1).

Fig. 1.

A systematic review was conducted to identify studies reporting structural or functional group differences between treatment-resistant depression (TRD) patients and healthy controls (HCs) or patients with non-treatment-resistant depression (non-TRD). Significant peak coordinates were extracted for each modality and entered into a coordinate-level coordinate network mapping (CNM) analysis, in which each coordinate was treated as an independent 6-mm spherical seed. Seed-to-whole-brain functional connectivity was estimated using resting-state fMRI data from 1000 HCs. Group-level T maps were corrected for family wise error (p < 0.05) binarized, and aggregated to generate voxel-wise probability maps. Voxels showing ≥50% convergence were retained to define TRD-associated networks. These networks were subsequently mapped onto canonical functional networks, assessed for cross-modality and cross-contrast correspondence, and examined for associations with neurotransmitter distribution maps.

Methods

Literature search and data extraction

We conducted a comprehensive search of PubMed, Scopus, and Web of Science for studies published up to August 15, 2025. To ensure completeness, we also screened the reference lists of relevant reviews. Studies were eligible if they reported voxel-wise, whole-brain analyses comparing clearly defined TRD patients with either HCs or non-TRD patients. Structural alterations had to be assessed using voxel-based, and functional alterations were required to capture localized intrinsic activity, such as regional homogeneity, amplitude of low-frequency fluctuations, or their derivative variants.

Study selection (according to predefined inclusion and exclusion criteria), coordinate extract and base information extract were conducted independently by two reviewers (Y.H. and H.Y.). Disagreements regarding eligibility or extracted information were resolved through discussion. The complete search strategy, screening criteria, and data-extraction procedures are detailed in the Supplementary Materials.

Coordinate network mapping

In contrast to traditional implementations that aggregate all coordinates from a study into a single mask, we applied a coordinate-based adaptation of the CNM framework. For every seed, we computed its connectivity profile within the normative connectome. This design preserves the anatomical specificity of individual loci, avoids study-level averaging that could mask heterogeneity, and improves the precision with which distributed functional networks are delineated.

Specifically, peak coordinates representing significant differences between TRD and either HC or non-TRD groups were extracted from all eligible studies. Each coordinate was modeled as an independent 6-mm spherical seed. Whole-brain functional connectivity (FC) for each seed was estimated using resting-state fMRI data from 1000 healthy adults in the Human Connectome Project [23]. Imaging data is preprocessed followed the procedures described in previous work [24]. For each seed, time series was obtained and correlated with every voxel in the brain to generate an individual FC map. These correlation coefficients were transformed to Fisher z-scores, and one-sample t-tests were applied across all participants to derive a group-level FC t-map for that seed. In line with standard CNM procedure, only positive connectivity was retained due to its clearer biological interpretability [25], given the ongoing debate regarding the biological interpretation of negative connectivity [26,27]. Furthermore, networks identified through anticorrelation mapping may not have a direct biological interpretation in the context of TRD in this study. Significant voxels were identified using a family-wise error (FWE) correction of p < 0.05, and each resulting t-map was binarized. All binarized seed maps corresponding to a given contrast were then aggregated to form a voxel-wise probability map, reflecting the proportion of seeds showing significant positive connectivity with each voxel.

To determine an appropriate network-overlap threshold, we performed a random ROI analysis in which the full CNM pipeline was repeated 10,000 times using randomly selected ROIs (matched in number to the real analysis) drawn from the Cole-Anticevic Brain-wide Network Partition (CAB-NP) atlas [22]. For each candidate threshold (40–55%), the 10,000 resulting random networks were combined to generate a null mask, where each voxel was labeled 1 if it appeared in any random network and 0 otherwise. Using the real ROI data, we then computed, for each threshold, the proportion of voxels in the real TRD-associated network that overlapped with the null mask. This proportion fell below 5% at the 50% threshold, which was therefore selected for all subsequent analyses. Further procedural details are provided in the Supplementary Text and Fig. S1–S2.

Sensitivity analysis

To assess the robustness of our findings, we performed a leave-one-out sensitivity analysis, in which one study was removed at a time from the CNM analysis to evaluate whether any individual study had a disproportionate influence on the results. Voxel-wise whole-brain spatial correlations were computed between the TRD-associated probability maps derived from each sensitivity analysis and those derived from all included studies. In addition, the Dice coefficients were calculated between the TRD-associated networks derived from each sensitivity analysis and those derived from all included studies. The Dice coefficient was defined as 2∗(number of overlap voxels)/(number of voxels in network A + number of voxels in network B). Spatial correlations and Dice coefficients were used to quantify the influence of each study on the findings. Studies showing relatively lower spatial correlation and Dice coefficient values were considered to have greater influence on the stability of the results.

Given that TRD samples in the literature are typically older, and because an age-matched validation analysis was not feasible, we stratified the healthy subjects into three age groups (<20 years, ≥20 and < 25 years, and ≥25 and ≤ 35 years) to examine the effect of age on the functional connectome underlying the identified networks, i.e., by performing a sensitivity analysis across different young-adult connectomes. CNM analyses were conducted separately for each age group. Voxel-wise whole-brain spatial correlations and Dice coefficients were then calculated to quantify the influence of age on the findings.

Validation analyses

To evaluate the robustness of the CNM results, we conducted two validation analyses. First, we repeated the full pipeline using alternative seed radii (4 mm and 8 mm) to assess spatial stability. Second, we reproduced the analysis in an independent test–retest subset of 747 participants (mean age = 21.08 ± 2.64 years; 338 males, 409 females) drawn from the same 1000 participants’ dataset using the original 6-mm radius, to confirm the reproducibility of the convergence patterns [23].

Overlap with canonical brain networks

For interpretability, we respectively quantified the spatial correspondence between the four TRD-associated networks and twelve canonical brain networks. Following Ji et al., the cortex and subcortex were parcellated into networks based on intrinsic connectivity organization, including the default mode (DMN), DAN, FPN, primary and secondary visual, cingulo-opercular (CON), somatomotor (SMN), language, auditory, ventral multimodal, posterior multimodal, and orbito-affective networks (Fig. S3) [22]. Spatial overlap was calculated as the proportion of shared voxels between each TRD-associated network and a canonical network, relative to the voxel count of the TRD-associated network. To complement this measure, we additionally quantified spatial overlap using the Dice coefficient, which captures the relative overlap between TRD-associated and canonical networks while accounting for the size of both networks.

Spatial correlation with neurotransmitter maps

To investigate the neurochemical substrates of TRD, we examined spatial correlations between TRD-associated probability maps and whole-brain distributions of neurotransmitter receptors and transporters. Nineteen in vivo molecular neuroimaging atlases covering nine major neurotransmitter systems were included (Table S1). Receptor and transporter densities were extracted from 116 regions defined by the Automated Anatomical Labeling (AAL) Atlas [28], yielding a 116 × 19 receptor/transporter profile matrix. We computed correlations between each probability map and neurotransmitter map.

To assess the significance of correlation, permutation testing was performed. Given the spatial autocorrelation inherent to both neurotransmitter and brain maps—where adjacent regions tend to exhibit similar values—the BrainSMASH toolbox was used to generate 1000 surrogate probability maps preserving the original spatial autocorrelation [29]. Correlations between these surrogates and neurotransmitter maps formed null distributions. Multiple comparisons were corrected using the Bonferroni method, with significance set at p < 0.002 (0.05/19 tests).

Furthermore, shared macroscale spatial gradients may influence the observed spatial correlations between TRD-associated probability maps and neurotransmitter distributions [30]. To account for this potential confound, we performed partial correlation analyses controlling for the principal cortical gradient. The first principal gradient (G1), which captures the dominant axis of cortical organization from unimodal to transmodal regions, was derived from Human Connectome Project data [30] and mapped onto the AAL atlas by averaging gradient values within each region. Partial correlations between TRD-associated probability maps and neurotransmitter maps were then computed while controlling for the regional G1 values. Statistical significance was assessed using the same permutation framework described above, which preserves spatial autocorrelation. The significance threshold was set at p < 0.002 (Bonferroni-corrected for 19 tests).

Cross-modality and cross-contrast network correspondence analysis

To characterize the correspondence between networks derived from different imaging modalities and contrasts, we quantified the spatial overlap and correlation among networks derived from different imaging modalities (structural vs. functional) and from distinct diagnostic contrasts (TRD–HC vs. TRD–non-TRD). First, substantive spatial overlap was evaluated by determining the proportion of voxels shared between two probability maps, expressed relative to the total voxel count within each network. Second, spatial correlations were computed between probability maps to quantify similarity in their spatial distributions. Given the strong spatial autocorrelation inherent in whole-brain maps, statistical significance was tested using a spatial-autocorrelation–preserving permutation test (as mentioned above). Multiple comparisons were controlled using Bonferroni correction.

Furthermore, Dice coefficients were calculated between the identified networks across modalities and contrasts as a complementary similarity metric.

Activation likelihood estimation (ALE) meta-analysis

All ALE analyses were conducted using GingerALE (version 3.0.2). ALE tests for spatial convergence against the null hypothesis that reported findings (i.e., whole-brain coordinates of case-control differences) are randomly distributed across the brain, rather than reflecting consistent abnormality in discrete brain regions. We performed ALE meta-analyses across all included articles to investigate spatial convergence in both TRD–HC and TRD–non-TRD contrasts, regardless of imaging modality (structural or functional) and direction of effect (hyper: patients > controls; hypo: patients < controls). Furthermore, for each contrast, we conducted separate ALE analyses for each modality, pooling reported abnormal brain regions irrespective of the direction of effects. Statistical significance was assessed using 5000 permutations with cluster-level inference (p < 0.05, family-wise error corrected), following recommended procedures for meta-analyses with 8–17 studies [31,32].

Results

Characteristics of included studies

From the initial pool of 9662 records, a total of 17 studies met inclusion criteria, yielding 15 TRD–HC contrasts (366 TRD patients and 360 HCs) and 10 TRD–non-TRD contrasts (295 TRD patients and 719 non-TRD patients) for the final analyses. The study selection procedure is detailed in Fig. S4. For each included study, key methodological and clinical characteristics, including authorship, publication year, sample size, demographic variables, diagnostic criteria, definition of TRD, and symptom severity, are summarized in Table S2–S3.

TRD-associated networks

Data derived from the TRD-HC contrast reported that the TRD-associated functional network predominantly encompassed the bilateral precentral gyrus, thalamus, insula, middle frontal gyrus (MFG), the left anterior cingulate cortex (ACC), superior frontal gyrus (SFG), and amygdala. In terms of canonical network composition, this network showed its greatest proportional overlap with the CON (48.4%), followed by the FPN (21.3%). The corresponding structural network included bilateral inferior frontal gyrus (IFG), insula, middle cingulate cortex (MCC), caudate, superior temporal gyrus (STG), MFG, amygdala, left middle temporal gyrus (MTG), putamen, pallidum, supramarginal gyrus, thalamus, and right Rolandic operculum, Heschl gyrus, and as well as the cerebellar crus II/I. This structural network exhibited its highest overlap with the CON (29.2%), FPN (23.6%), DMN (13.3%), and LN (12.9%) (Fig. 2 and Table S4).

Fig. 2.

Part (A) presents TRD-associated functional and structural networks generated by coordinate network mapping analysis, with seeds derived from peak coordinates identified in the TRD–HC and TRD–non-TRD contrasts, respectively. Part (B) shows a stacked bar chart summarizing the cumulative spatial overlap of the four reconstructed networks with canonical functional networks. Part (C) displays radar charts summarizing the proportional overlap between each TRD-associated network and the canonical functional networks. TRD = treatment-resistant depression; non-TRD = non-treatment-resistant depression; HC = healthy control.

For the TRD–non-TRD comparison, the TRD-associated functional network included the bilateral precentral gyrus, insula, supramarginal gyrus, IFG, supplementary motor area (SMA), MCC, Rolandic operculum, MFG, right precuneus, left fusiform and ACC. This network overlapped most prominently with the CON (65.4%) and FPN (8.4%). The structural network for this contrast involved bilateral MCC, insula, STG, putamen, inferior temporal gyrus (ITG), MTG, IFG, fusiform gyrus, ACC, SMA, amygdala, parahippocampel gyrus, Rolandic operculum, precentral gyrus, right pallidum, left supramarginal gyrus, and MFG. Its dominant canonical associations were with the CON (44.7%), DAN (8.7%), and DMN (8.2%) (Fig. 2 and Table S4).

Dice coefficient analyses were consistent with the spatial overlap results, with values highest for the CON across all four TRD-associated networks (Table S5).

Cross-modality and cross-contrast network correspondence analysis

The spatial overlap between functional and structural networks was modest. In the TRD–HC contrast, only 10.3% of the functional network overlapped with 1.5% of the structural network. In the TRD–non-TRD contrast, 20.7% of the functional network overlapped with 15.8% of the structural network. For TRD-associated functional networks, 42.6% of the TRD–HC network intersected with 5.5% of the TRD–non-TRD network. For TRD-associated structural networks, 20.9% of the TRD–HC network overlapped with 14.1% of the TRD–non-TRD network. In addition, using the Dice coefficient as a complementary similarity metric, values were also relatively low across all comparisons (cross-modality in TRD–HC: 0.03; cross-modality in TRD–non-TRD: 0.18; cross-contrast in the functional modality: 0.10; cross-contrast in the structural modality: 0.17).

Voxel-wise spatial correlations revealed strong cross-modal and cross-contrast correspondence. Functional and structural networks were strongly correlated in both TRD–HC (ρ = 0.66) and TRD–non-TRD (ρ = 0.57) comparisons. Spatial similarity across contrasts was also high, with ρ = 0.69 between the two functional networks and ρ = 0.76 between the two structural networks. All correlations remained significant after spatial-autocorrelation–preserving permutation testing (p < 0.001) (Fig. 3).

Fig. 3.

Scatter plots and corresponding permutation tests illustrating spatial correlations of TRD-associated networks, examining both cross-modality and cross-contrast consistency. Cross-modality correlations were assessed between functional and structural networks derived from the TRD–HC contrast and from the TRD–non-TRD contrast, respectively. Cross-contrast correlations were evaluated the consistency of structural networks and of functional networks across the two contrasts. TRD = treatment-resistant depression; non-TRD = non-treatment-resistant depression; HC = healthy control.

Spatial correlation with neurotransmitter maps

In the TRD–HC contrast, the functional probability map showed significant spatial correlations with the N-Methyl-d-aspartate receptor (NMDA, r = 0.53) and the dopamine transporter (DAT, r = 0.47). The structural probability map demonstrated a similar pattern, correlating strongly with NMDA (r = 0.49), DAT (r = 0.43), and metabotropic glutamate receptor 5 (mGluR₅, r = 0.63). In the TRD–non-TRD contrast, the functional probability map correlated with several neurotransmitter systems, including the 5-Hydroxytryptamine Receptor 1B (5-HT_1B, r = 0.62), 5-HT₆ receptor (5-HT6, r = 0.62), mGluR₅ (r = 0.58), norepinephrine transporter (NET, r = 0.54), and NMDA (r = 0.44). The structural probability map showed significant correlations with DAT (r = 0.41) and NMDA (r = 0.39) (Fig. 4). All of these associations were significant in spatial-autocorrelation–preserving permutation testing (p < 0.001) (Fig. S5). Additional results are presented in Fig. S6.

Fig. 4.

Spatial correlations between neurotransmitter maps and TRD-associated probability map. Schematic representations depict the spatial distributions of individual neurotransmitter receptors or transporters, and radar charts summarize the correlation patterns between the probability maps and each neurotransmitter map. The left subfigure presents the TRD vs. HC contrast, and the right subfigure presents the TRD vs. non-TRD contrast. 5-HT = 5-hydroxytryptamine; 5-HT1A/1B/2A/4/6 = 5-HT receptor subtypes; 5-HTT = serotonin transporter; α4β2 = nicotinic acetylcholine receptor subtype α4β2; CB1 = cannabinoid receptor type 1; D1/D2 = dopamine receptors; DAT = dopamine transporter; GABAa = gamma-aminobutyric acid A receptor subtype; mGluR₅ = metabotropic glutamate receptor 5; MOR = mu-opioid receptor; M1 = muscarinic acetylcholine receptor subtype 1; VAChT = vesicular acetylcholine transporter; H3 = histamine receptor subtype H3; NET = norepinephrine transporter; TRD = treatment-resistant depression; non-TRD = non-treatment-resistant depression; HC = healthy control.

In the partial correlation analyses controlling for the principal cortical gradient, most of these associations remained significant in spatial-autocorrelation–preserving permutation testing (all p < 0.002) (Fig. S7). However, in the TRD–non-TRD contrast, the functional probability map was no longer significantly correlated with NMDA distribution, and the structural probability map was no longer significantly correlated with DAT distribution.

Sensitivity analysis

In the leave-one-out sensitivity analysis (Table S6), the mean Dice coefficients in the TRD–non-TRD contrast were 0.66 (standard deviation [SD] = 0.21; range: 0.35–0.92) for the functional modality and 0.64 (SD = 0.19; range: 0.43–0.80) for the structural modality. Relatively large influences were observed for the studies of Sun et al. [33] and Amen et al. [34] in the functional modality (Dice coefficients = 0.35 and 0.42 after exclusion, respectively) and for the study of Sandu et al. [35] in the structural modality (Dice coefficient = 0.40 after exclusion). In the TRD–HC contrast, the mean Dice coefficients were 0.65 (SD = 0.10; range: 0.46–0.83) for the functional modality and 0.63 (SD = 0.11; range: 0.43–0.72) for the structural modality. Relatively large influences were observed for the study of Li et al. [36] in the functional modality (Dice coefficient = 0.46 after exclusion) and for the study of Johnston et al. [37] in the structural modality (Dice coefficient = 0.43 after exclusion). In addition, the TRD-associated probability maps derived from each sensitivity analysis were significantly positively correlated with those derived from all included studies (p < 0.001, 10,000 iterations, correlation coefficient = 0.54–0.99).

In the age-stratified sensitivity analysis of the functional connectome, although the TRD-associated probability maps derived from different age groups were significantly positively correlated with those derived from all subjects (p < 0.001, 10,000 iterations), the Dice coefficients between the TRD-associated networks derived from each age group and those derived from all subjects were modest (Table S7).

Validation analyses

Findings from the retest dataset exhibited strong concordance with those from the test dataset, with only minor discrepancies likely attributable to difference in sample size (Fig. S8). Furthermore, repeating the CNM construction procedure using alternative seed sizes (4 mm and 8 mm radii) yielded networks highly consistent with those derived from the 6 mm radius, indicating that the findings were robust to variations in seed size (Fig. S9–S10).

ALE meta-analysis

In the TRD–HC contrast, regardless of imaging modality and direction of effect, one significant cluster was identified in the precuneus (Montreal Neurological Institute (MNI): 6, −56, 30; cluster size = 82 voxels; ALE = 0.02; Z value = 4.37; Fig. S11). In the TRD–non-TRD contrast, regardless of imaging modality and direction of effect, one significant cluster was observed in the superior temporal gyrus extending to the supramarginal gyrus (MNI: 48, −42, 22; cluster size = 94 voxels; ALE = 0.02; Z value = 4.60; Fig. S11). However, when analyses were conducted separately for each modality, no significant clusters were identified in either contrast.

Discussion

This CNM study provides a comprehensive multimodal characterization of TRD by incorporating both TRD–HC and TRD–non-TRD contrasts. The network-level organization of TRD pathology showed a highly convergent pattern: the general burden of illness was mapped predominantly onto the CON and FPN, whereas abnormalities specifically associated with treatment resistance were more selectively localized to the CON. Validation analyses demonstrated that these convergence patterns were highly reproducible across test–retest samples and remained stable under variations in seed size. High Dice coefficients and significant spatial correlations in leave-one-out sensitivity analyses also support the robustness of these findings. Spatial correlation analyses further anchored these network-level abnormalities to glutamatergic, dopaminergic, and monoaminergic neurotransmitter systems. Together, these observations provide a coherent context for the more detailed interpretation of TRD-associated network and neurochemical alterations presented below.

In the TRD-HC contrast, convergent functional and structural networks were predominantly located within the CON and FPN. These two networks constitute a dual top-down control architecture: the FPN supports flexible, moment-to-moment adjustments of cognitive control, whereas the CON sustains stable task-set maintenance and integrates performance feedback [38,39]. Aberrant connectivity among key nodes, such as the subgenual cingulate, ACC, striatum, and dlPFC, has been reported, implicating dysregulation of the two networks in depression [[40], [41], [42], [43]]. At the network level, depression is characterized by disrupted intrinsic connectivity within the CON and FPN, disrupted inter-network connectivity between them, and altered topographic organization affecting their territory [[44], [45], [46], [47]]. Such disruptions may bias the inappropriate attribution of salience to external cues, contributing to heightened threat sensitivity and persistent rumination [[47], [48], [49], [50]]. The cross-modality convergence observed in the present study indicates that CON and FPN represent robust, architecture-level features of TRD.

In contrast, neuroimaging differences between TRD and non-TRD patients were selectively connected to the CON, highlighting this network as the key distinguishing feature of treatment resistance. The CON spans widespread cortical and subcortical regions, with the anterior insula and dorsal ACC as hubs, and includes the MCC, SMA, precuneus, SFG, MFG, putamen, pallidum, precentral gyrus, Rolandic operculum, and supramarginal gyrus [22]. Functionally, the CON supports tonic alertness and the rapid detection of behaviorally relevant information: the anterior insula acts as an alerting hub that engages the dorsal ACC, supramarginal gyrus, and lateral prefrontal regions to update internal models and adjust control states [51]. Given its central role in cognitive control and affective regulation, CON dysfunction may represent a core vulnerability underlying treatment resistance, although it remains unclear whether this dysfunction constitutes a premorbid trait or arises secondarily with illness progression. Evidence from neuromodulation studies shows that altering connectivity of the anterior insula and ACC can produce symptomatic improvements in TRD [52,53]. These findings still underscore the centrality of CON circuitry in the pathophysiology of TRD.

The ALE meta-analyses identified only one significant cluster in each contrast, regardless of imaging modality and direction of effect, indicating limited spatial convergence of reported abnormalities across included studies. Furthermore, when analyses were conducted separately for each modality, no significant clusters were identified in either contrast. These ALE findings differ from those obtained using CNM. This discrepancy likely reflects fundamental differences in methodological logic: ALE is designed to identify consistent spatial convergence of reported abnormalities across studies [54], whereas CNM aims to identify brain networks functionally connected to these distributed abnormalities, thereby capturing the network-level substrates underlying the disorder [18,55]. Importantly, CNM is based on the assumption that, although reported abnormalities may appear spatially heterogeneous across studies, they may converge onto a common network that contributes to the manifestation of the disorder [18,55]. Furthermore, psychiatric disorders, including depression, are increasingly conceptualized as disorders of network dysfunction rather than focal brain region abnormalities [56]; thus, CNM may provide greater sensitivity than conventional coordinate-based meta-analytic approaches in identifying the neural substrates underlying psychiatric disorders.

Two perspectives have been proposed regarding the neurobiology of TRD. One conceptualizes TRD as a continuum of depression, with treatment resistance representing a more severe expression of a shared underlying pathophysiology. The other considers TRD as a distinct subtype with qualitatively different mechanisms [5]. Our cross-contrast findings, in both functional and structural modalities, tend to favor the former interpretation, as the TRD-associated probability maps derived from the TRD–HC and TRD–non-TRD contrasts showed a high degree of spatial correspondence. Collectively, these findings suggest that TRD shares core network-level alterations with broader MDD pathology, while CON dysfunction may represent a neural feature associated with treatment resistance. However, this interpretation should be made with caution, as the current design does not allow a direct dissociation between general disease burden and treatment resistance. Differences in comparator groups, TRD definitions, illness characteristics, and treatment histories across studies may all contribute to the observed patterns. Therefore, the present findings should be interpreted as suggestive rather than definitive.

Some regions were more frequently identified in the left hemisphere within the TRD-related networks, although the overall network distribution remained largely bilateral. Cerebral lateralization refers to the specialization of certain neural functions to one hemisphere, which possibly of importance for social interactions and cognitions [57]. Consistent with this, a recent DIRECT consortium study reported lateralization in the DMN, attention network and control network in patients with MDD [58]. Nevertheless, as lateralization was not formally examined in the present study, these observations should be interpreted cautiously.

Notably, high spatial correlations but relatively low voxel-wise overlap were observed in cross-modality and cross-contrast correspondence analyses. This discrepancy reflects the distinct properties captured by different similarity metrics: spatial correlation quantifies the similarity of continuous spatial distributions, whereas overlap-based measures (e.g., the Dice coefficient) depend on thresholded maps and are sensitive to boundary definitions. As a result, two maps may exhibit highly similar global spatial patterns while showing limited voxel-wise overlap after thresholding. For whole-brain mappings, different thresholds can yield different network configurations. As a nonlinear operation, thresholding continuous data discards weak yet consistent signals and amplifies boundary effects [59]. Additionally, variability observed in sensitivity analyses suggests that the limited number of included studies may further contribute to instability in voxel-level overlap. Overall, the discrepancy between spatial patterns and voxel-wise overlap primarily reflects the effects of thresholding and metric-specific sensitivity, rather than a true inconsistency in the underlying spatial organization. Future studies may mitigate this issue by including more original studies and employing complementary similarity metrics to provide a more robust and comprehensive assessment of network correspondence.

As is well established, the efficacy of neuromodulatory interventions such as DBS critically depends on the connectivity profile of the stimulation site with other brain regions [60]. Accordingly, this connection-based principle aligns with the conceptual basis of CNM, emphasizing that the TRD-associated network identified in our analysis may represent a biologically meaningful intervention target [19]. Importantly, when the dlPFC is selected as a stimulation target, clinical improvements have been found to depend on its connectivity with the anterior insula, subgenual cingulate, and striatum, and modulation of these connections reliably predicts treatment response [43,61,62]. These observations suggest that the regions traditionally regarded as “connected” to the stimulation site may in fact constitute the true therapeutic substrates. Consequently, direct modulation of the nodes within the TRD-associated network identified here may provide enhanced therapeutic leverage. Supporting this notion, previous studies have shown that networks constructed from patient–control lesion differences and those from neuromodulatory intervention sites converge on overlapping brain networks, providing evidence for causal circuitry underlying depression [63]. Together, these findings suggest that the TRD-associated network identified here captures biologically meaningful circuitry relevant to both disease expression and therapeutic modulation.

The pathology of TRD likely extends beyond alterations in a single neurotransmitter or reduced transporter availability, instead reflecting a complex imbalance and redistribution across multiple neurotransmitter systems. This may partly explain why increasing the dose of conventional serotonergic medications often fails to improve clinical outcomes and may even exacerbate adverse effects [64,65]. Our findings highlight the potential involvement of glutamatergic signaling, particularly NMDA and mGluR₅ receptors, in the neuropathology of TRD. As the primary excitatory neurotransmitter in the central nervous system, glutamate plays an important role in neurodevelopment and cognitive function [66]. The observed positive spatial correlations suggest that brain regions with higher glutamate receptor distributions tend to show stronger associations with TRD, potentially indicating a selective vulnerability of glutamatergic systems in TRD. This interpretation is further supported by proton magnetic resonance spectroscopy studies showing reduced glutamine levels in the medial frontal cortex in depression, a region also involved in all four TRD-associated networks [67]. In parallel, pharmacological modulation of glutamatergic neurotransmission, particularly via NMDA and mGluR₅ receptor antagonism, has emerged as a promising therapeutic strategy for TRD [68]. In addition to glutamatergic findings, we observed significant spatial correspondence between TRD-associated probability maps and the normative distribution of the DAT, particularly in the TRD–HC contrast across both functional and structural modalities. These findings suggest that TRD-associated networks preferentially involve brain regions with relatively high DAT expression. Importantly, this spatial correspondence does not imply increased DAT availability in TRD, but rather indicates a potential relationship between the topography of TRD-associated networks and the organization of the dopaminergic system. Regions within the identified network, particularly the striatum, have previously been reported to show lower DAT availability in MDD compared with HCs [69], suggesting altered dopaminergic neurotransmission in depressive disorders. Our findings are broadly consistent with this literature in implicating the dopaminergic system in TRD. However, the two approaches address different biological levels: positron emission tomography (PET) and single-photon emission computed tomography (SPECT) studies directly measure DAT availability, whereas the present spatial correlation analysis examines whether TRD-associated networks are preferentially distributed in regions enriched for DAT expression. Dopaminergic dysfunction has long been implicated in anhedonia, motivational deficits, and reward-processing abnormalities in depression [70,71]. Therefore, the observed association between TRD-associated networks and DAT distribution may reflect the involvement of dopaminergic circuits in the network-level pathophysiology of TRD. Nevertheless, the spatial correlations with neurotransmitter maps remain correlational and do not establish direct mechanistic involvement of specific neurotransmitter systems in TRD. Overall, these findings suggest that TRD-associated networks are unlikely to be randomly distributed, but instead preferentially involve glutamatergic and dopaminergic systems associated with cognition, reward processing, and motivation.

There are several limitations that should be acknowledged. First, although TRD is generally defined as failure of at least two adequate antidepressant trials, study-specific inclusion criteria remain heterogeneous. Specifically, variations exist in how “adequate treatment trials” are defined (e.g., whether limited to pharmacotherapy alone or extended to include physiotherapy) as well as in the required duration for pharmacotherapy (e.g., 4, 6, or 8 weeks), resulting in variability in sample characterization. Moreover, TRD frequently co-occurs with other psychiatric conditions or substance-use behaviors, further obscuring disorder-specific pathophysiology [72]. Compounding these issues, patients with TRD have been exposed to multiple treatment modalities, and different treatment can induce distinct morphometric and functional brain changes [73]. Second, imaging-related parameters varied across the included studies. Differences in magnetic field strength, slice thickness, preprocessing pipelines, and statistical methods may have influenced the results and could potentially affect the robustness of our conclusions.

Third, the cross-sectional design and modest sample sizes of the included neuroimaging studies limit causal inference and constrain the generalizability of our findings. To date, few studies have prospectively examined brain alterations from illness onset through the development of treatment resistance. To identify what biologically underlies treatment resistance, future work should extend CNM analyses in three complementary directions: (1) mapping whether CNM-derived networks differ between responders and non-responders during the first antidepressant trial in drug-naïve MDD patients; and (2) mapping whether CNM-derived networks differ between patients with MDD who have not yet reached treatment resistance but show differential treatment outcomes in antidepressant trial.

Forth, the CNM approach treats all original study results equally, which may dilute effect specificity. Future work could incorporate more detailed information from original reports, such as empirically derived clusters or effect-size–weighted regions of interest, rather than uniformly applying standardized spherical seeds. Moreover, the CNM framework depends on normative resting-state functional connectomes from healthy participants. These normative connectomes may not fully capture disease-related reconfiguration of network topology. Furthermore, although significant spatial correlations were observed, the Dice coefficients were modest in the sensitivity analyses stratified by connectome age. While the reduced sample size in these sensitivity analyses may contribute to this effect, the findings also suggest that connectome age may influence convergence patterns, consistent with prior evidence that functional connectomes exhibit dynamic changes across the lifespan [71,74,75]. Future studies incorporating patient-specific and age-matched connectomes may provide more accurate mapping of TRD-associated network.

Our CNM analysis demonstrates that heterogeneous neuroimaging abnormalities reported across individual TRD studies converge onto a common large-scale control architecture centered on the CON and FPN, with CON function emerging as a specific neural signature of treatment resistance. The robustness of these findings is supported by validation analyses and cross-modality correlations. Together, these results indicate that the TRD-associated network identified here captures biologically meaningful circuitry linking disease expression with therapeutic modulation.

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Author contributions

Y.H. and H.Y. conceived the study and performed the literature search and study selection. Y.H. and H.Y. extracted and curated the data. Y.H. conducted the formal analyses and prepared the figures and tables. H.Y. verified the extracted data and checked the analytical results. H.C. and W.G. provided methodological guidance and supervised the study. W.G. contributed to formal analysis and validation of the results. Y.H. drafted the original manuscript. Y.H., H.Y., and W.G. revised and edited the manuscript. H.C. and W.G. acquired funding for the study. All authors reviewed and approved the final manuscript.

Funding sources

This study was supported by grants from the NIH (R01MH138682 and R01MH137501), the National Natural Science Foundation of China (Grant No. 82171508).

Declaration of competing interest

The authors report no biomedical financial interests or potential conflicts of interest.

Appendix A. Supplementary data

The following is the Supplementary data to this article: