Structural Brain Network Alterations in Relation to Treatment and Illness Severity in Bipolar Disorder

Abstract

Background:

Large-scale T1-weighted MRI studies have established grey-matter abnormalities in bipolar disorder (BD), with our group contributing to consensus findings. However, structural connectivity, particularly within emotion- and reward-related circuits, remains poorly understood. Diffusion-weighted MRI (dMRI) enables investigation of white-matter pathways, yet prior work is constrained by small samples, methodological heterogeneity, and unclear medication effects. We conducted the largest dMRI network analysis in BD, relating symptom burden and polypharmacy to tractography-derived connectivity and graph-theoretic metrics.

Methods:

Cross-sectional structural and diffusion MRI scans from 449 individuals with BD (35.7±12.6 years) and 510 controls (33.3±12.6 years), aged 18–65, were analyzed across 16 ENIGMA-BD sites. Standardized segmentation/parcellation and constrained spherical deconvolution tractography generated individual structural connectivity matrices. Graph-theoretic metrics of global and subnetwork organization were related to symptom severity and medications.

Results:

BD showed widespread network alterations (lower density and efficiency, longer path length, and higher betweenness centrality), altered microstructural organization in a limbic–basal ganglia circuit, and abnormal streamline counts in a default-mode/salience/fronto-limbic–basal ganglia network. Longer illness duration, later onset, and psychosis history were associated with greater abnormalities in network architecture, whereas more manic episodes were associated with greater fronto-limbic connectivity. Antidepressant (particularly SSRI), anticonvulsant, and antipsychotic use related to poorer global and fronto-limbic connectivity; no clear lithium effects emerged.

Conclusions:

As the largest structural connectivity study in BD, we reveal widespread disruption in reward and emotion-regulation networks influenced by illness severity and medication use. Results show that multisite harmonization is feasible and highlight ENIGMA-BD as a scalable framework for identifying reproducible neurobiological markers.

Introduction

Bipolar disorder (BD) is a chronic, recurrent mood disorder characterized by depression and (hypo)mania, affecting roughly one in 25 adults in the United States and contributing substantially to global disability (1). Although cognitive and affective impairments are well-documented, treatment response remains highly variable: only ~30% of patients respond robustly to lithium–the standard first-line agent (2)–and many continue to experience persistent symptoms, functional impairment, or treatment resistance (3,4). BD also carries one of the highest suicide rates among psychiatric disorders (5). Mapping neural circuit dysfunction is therefore critical for understanding illness burden, predicting treatment response, and advancing biologically grounded, personalized strategies.

BD is characterized by dysregulation within fronto-limbic circuits and disrupted interactions with networks supporting fear, anxiety, and cognitive control (6,7). Diffusion-weighted MRI studies show that white matter (WM) abnormalities extend beyond anterior fronto-limbic pathways to major association and projection fibers, including the superior longitudinal (arcuate), inferior longitudinal, inferior fronto-occipital fasciculi, posterior thalamic radiations, and internal capsule (8,9). Robust effects are also observed in the corpus callosum and cingulum, where severity relates to age of onset and illness duration (8), although the contribution of callosal transfer to mood switching and emotional regulation remains unclear (10,11). Altered connectivity in posterior and cerebellar pathways may reflect compensation for disrupted frontal circuits (12). WM disruptions are particularly pronounced in psychotic BD presentations (13) and may contribute to impaired cognitive control and affective instability.

Network neuroscience reframes brain function from regional localization to connectivity-based models, emphasizing distributed interactions among neural systems. Graph theory provides a mathematical framework for this approach, representing regions as nodes and white-matter pathways as edges, enabling quantification of network organization and integration (14). Although diffusion and structural imaging studies report focal WM abnormalities in BD (15), prior graph-theoretic work has relied on small samples with heterogeneous pipelines, leading to inconsistent network findings. Variability in acquisition, segmentation, parcellation, and connectivity metrics, as well as differences in illness duration, severity, and medication exposure, further contributes to these discrepancies. Large-scale, harmonized network analyses that control for site effects are needed to generate more reproducible estimates of BD circuitry, clarify illness-related variation, and determine how common treatments shape network architecture.

Most patients with BD require concurrent use of multiple medications (polypharmacy) to manage symptoms. Large multi-site studies from The Enhancing Neuro Imaging Genetics through Meta-Analysis Bipolar Disorder Working Group (ENIGMA-BD) have reported associations between psychotropic medications and structural brain features, linking lithium to greater cortical thickness in parietal and frontal regions, and anticonvulsants and antipsychotics to thinner cortex and lower surface area in occipital and frontal regions (16,17). In the first ENIGMA-BD diffusion-weighted MRI (dMRI) study (N=3,033; 26 sites), Favre et al. (2019) (8) found, using diffusion tensor imaging (DTI), that lithium associated with higher fractional anisotropy (FA) and lower mean diffusivity (MD), while antipsychotics and anticonvulsants linked to lower FA. Similar analyses using advanced WM connectivity features such as graph theory are needed to clarify relationships between network dysfunction, treatment, and illness severity in BD.

ENIGMA-BD is the largest global consortium for harmonized neuroimaging and clinical data in BD, enabling standardized processing and improved replicability of brain signatures (18). Using a standardized pipeline across 16 ENIGMA-BD datasets (Figure S1), we assessed whole-brain and subnetwork structural connectivity in BD versus controls. Structural connectivity can be quantified using complementary weighting schemes that capture different aspects of WM organization. FA-weighted networks reflect microstructural properties of white matter pathways, such as axonal coherence and myelination, whereas networks weighted by the number of streamlines (NOS) represent tractography-derived estimates of connection density (19,20). Each metric provides complementary insights into microstructure organizational properties as they are arranged in large-scale architectural patterns of connectivity of structural brain networks (20). We hypothesized that BD would show network-level disruptions: reduced clustering and efficiency, longer characteristic paths, and higher betweenness centrality, alongside alterations in limbic, basal ganglia, interhemispheric, and cerebellar projections (8,12,21). We further predicted that greater symptom burden would relate to more pronounced abnormalities (8), and that medication exposure, particularly lithium, anticonvulsants, and antipsychotics (especially when categorized by pharmacological mechanisms of action) would account for additional variance in WM network organization, consistent with prior ENIGMA findings (8,16,17). Overall, our goal was to examine the relationships between network metrics, treatment responses, and symptoms, to advance the current understanding of BD-related structural networks and identify biologically grounded targets for interventions.

Methods and Materials

Subjects

Sixteen independent ENIGMA-BD cohorts contributed 959 participants (449 BD, 510 controls; 51.1% female), aged 18–65, with DSM-IV-R diagnoses of BD type I or II confirmed by clinical interview with a psychiatrist (demographics and clinical details in Tables 1 and S2). Individuals with BD type I (BD-I N=201) and type II (BD-II N=42) were included in the analysis where diagnostic subtype information was available. For several cohorts, subtype classification was not consistently reported in the harmonized dataset, resulting in a proportion of BD participants without subtype specification. These individuals were retained in the primary analyses to maximize sample size, and follow-up tests were conducted in the subset with available subtype information to examine BD-I and BD-II differences. Analyses focused on adults (18–65), an age range commonly used in prior studies and best represented across sites; participants younger than 18 or older than 65, or with poor-quality MRI data after inspection were excluded from our analyses. Site-specific inclusion/exclusion criteria are detailed in Table S3. All participating sites obtained approval from their local institutional review boards and ethics committees, and all study participants provided written informed consent.

Table 1.

Clinical and Sociodemographic Details of Participants.

	Controls N = 509	Bipolar Disorder N = 450	Statistical Comparison Diagnostic Groups
Age, Mean (SD)	33.25 (12.61)	35.65 (12.59)	t943 = 2.9, p = 0.003
Sex, n (%)			χ21 = 6.9, p = 0.008
Male	228 (45%)	241 (54%)
Female	281 (55%)	209 (46%)

n/N = Sample/Total Sample Size; SD = Standard Deviation. t = t-test. χ²= chi-square.

Image acquisition and processing

Contributing sites shared raw T1-weighted and diffusion-weighted (dMRI) scans with the central processing team (acquisition parameters in Table S4). Subject-specific cortico-subcortical brain networks were mapped probabilistically across 86 regions – 34 cortical and 9 subcortical bilaterally plus cerebellum – using FreeSurfer v5.3.0 (22), with the Desikan-Killiany atlas (23) applied to the T1-weighted images (Figure 1). Diffusion MR images were processed using deterministic constrained spherical deconvolution (CSD) to resolve crossing fibers (ExploreDTI v4.8.6) (24,25), as described in Nabulsi et al. (2019) (21) and detailed in Supplementary Material. Structural connectivity matrices (86×86) were weighted by fractional anisotropy (FA) and streamline count (NOS); binary matrices were also derived. Network topology was examined across BD and control groups using FA- and NOS-weighted and unweighted matrices.

Figure 1. Human Connectome Reconstruction.

This figure illustrates the standardized process of human connectome reconstruction and subsequent analyses. Proceeding from left to right: Brain Region Definition (node): The brain is divided into 34 cortical and 9 subcortical regions bilaterally using FreeSurfer v5.3.0, based on the Desikan-Killiany atlas. White Matter Trajectory Reconstruction (edge): White matter pathways are reconstructed using a deterministic non-tensor constrained spherical deconvolution algorithm, implemented via ExploreDTI v4.8.6. Brain Network Construction: The resulting brain network, composed of nodes and edges (human connectome) is represented as a structural connectivity matrix. Matrices are both unweighted and weighted by fractional anisotropy or the number of streamlines, generated using ExploreDTI v4.8.6. Global Network Measures: From the structural connectivity matrices, whole-brain network measures describing features of segregation and integration are calculated using the Brain Connectivity Toolbox v1.52). Linear mixed models (LMMs) are fitted to investigate structural connectivity differences and the influence of medications and clinical variables; a permutation analysis is performed to identify subnetwork-level differences between patients and controls, using NBS v1.2 (Zalesky et al. 2010).

Whole-brain measures derived from the connectome

Global parameters summarizing whole-brain connectivity properties were extracted from both unweighted and weighted matrices (Figure 1). Measures of segregation, including global density and clustering coefficient were computed as the mean of the respective 86 regional estimates using functions from the Brain Connectivity Toolbox v1.52 (26). Measures of integration such as characteristic path length and efficiency were derived, as well as measures of influence and centrality, global degree/strength and betweenness (detailed in Table S5).

Statistical analysis of the structural connectome

Linear mixed models (LMMs) assessed relationship between whole-brain normally distributed measures and diagnosis, while adjusting for age, sex and site. In this model, diagnosis, age, and sex were included as fixed effects, and site was included as a random intercept to account for between-site variability (R v4.2.1) (Figure 1). Age-by-diagnosis and sex-by-diagnosis interactions were also tested. False Discovery Rate (FDR) correction (p_FDR<.05) (27) was applied across all tests, consistent with prior graph theory work (19,27). Network-Based Statistics (NBS v1.2) (26) performed mass univariate testing on FA- and NOS-weighted connectivity graphs (Figure 1), identifying subgraphs with significantly weaker or stronger connectivity while controlling for family-wise error rate (FWER). An F-test adjusting for age, sex, and site tested group connectivity differences (M=5000; p<.05); connections were thresholded to retain only suprathreshold edges, with FWER correction via permutation testing ensuring validity regardless of threshold choice (26).

Clinical associations in bipolar disorder

Connectivity associations with clinical variables were tested in the BD group using linear mixed-effects models (p_FDR<.05). For illness duration and age at onset, age was residualized to isolate effects from age-related influences. Polypharmacy was examined under two frameworks: (1) conventional indication-based categories (lithium, antipsychotics, antidepressants, anticonvulsants; Table S1), and (2) Neuroscience-Based Nomenclature (NbN) (30), which classifies medications by neurobiological targets, using medication names or Anatomical Therapeutic Chemical (ATC) codes (Supplement Table S6). Primary analyses focused on illness duration, age at onset, and lithium, anticonvulsant, and antipsychotic use; exploratory models included symptom-severity variables (e.g., psychosis history, episode counts) and NbN categories. All clinical models tested whether effects persisted when controlling for medication use, while medication models adjusted for concurrent medications (polypharmacy) and illness-course severity (the number of manic and depressive episodes), ensuring distinct contributions of clinical and pharmacological factors. Model specifications are provided in Table S7.

Results

Participants’ clinical and demographic characteristics

Individuals with BD were predominantly euthymic (86%) at the time of scanning, significantly older than controls and comprised a higher proportion of females (55%) (Table 1). Clinical and sociodemographic details of BD are provided in Table S8.

Whole-brain measures of network integration and segregation

BD and controls significantly differed in integration and segregation connectivity metrics (Figure 2; Table 2). Unweighted network topology in BD showed lower density, longer characteristic path length, lower global efficiency, and higher betweenness centrality (Cohen’s d=−.2–.2; p_FDR=[.01-.04]). Similar disruptions were observed in NOS-weighted networks, with longer path length and lower efficiency in BD (d=.2; p_FDR=.001), while FA-weighted networks showed no group differences. Males showed lower clustering and local efficiency (d=−.3; p_FDR=.0002) than females in NOS-weighted networks, independent of diagnosis, with no sex-by-diagnosis interaction. Age was significantly associated with lower density and efficiency (unweighted), and with path length, local efficiency, and betweenness centrality (NOS-weighted) (p_FDR=[.007-.03]), but no age-by-diagnosis interaction was detected (p_FDR=[.8-.9]). Sorted bar plot of Cohen’s d effect sizes for whole-brain network metrics in Figure S2.

Figure 2. Whole-brain measures affected in bipolar disorder.

Global network measures affected in BD, relative to controls. The BD group exhibited dysconnectivity compared to healthy controls across unweighted and NOS-weighted networks. Dysconnectivity was defined by lower global density, longer path length and lower global efficiency; bars represent mean±SD, raw means are plotted. NOS = number of streamlines. *p_FDR <.05

Table 2.

Global Network Measures Across Unweighted and Weighted Networks.

Network Metric, Mean±SD	Healthy Controls N = 509	Bipolar Disorder N = 450	pFDR-value	Cohen's d
Binary Density	0.354±0.082	0.349±0.086	0.018*	0.19
Binary Degree	48.685±37.111	48.690±37.222	0.666	−0.04
Binary Clustering Coefficient	0.652±0.036	0.648±0.045	0.786	0.03
Binary Characteristic Path Length	1.527±0.567	1.541±0.568	0.018*	−0.20
Binary Global Efficiency	0.666±0.049	0.662±0.050	0.021*	0.18
Binary Local Efficiency	0.927±0.008	0.927±0.008	0.067	0.15
Binary Betweenness Centrality	60.461±10.763	61.450±11.265	0.042*	−0.16
FA Strength	12.323±9.527	12.094±9.486	0.121	0.12
FA Clustering Coefficient	0.247±0.165	0.246±0.163	0.138	0.11
FA Characteristic Path Length	4.810±1.441	4.861±1.430	0.397	−0.07
FA Global Efficiency	0.288±0.162	0.281±0.162	0.946	0.01
FA Local Efficiency	0.345±0.220	0.347±0.218	0.367	0.08
FA Betweenness Centrality	80.308±20.387	81.264±19.584	0.369	−0.07
NOS Strength	2,728.270±981.834	2,583.344±1,014.681	0.090	0.13
NOS Clustering Coefficient	35.332±8.302	33.635±7.587	0.987	0.00
NOS Characteristic Path Length	0.011±0.004	0.012±0.004	0.018*	−0.20
NOS Global Efficiency	125.679±50.367	119.366±54.184	0.018*	0.21
NOS Local Efficiency	62.883±14.794	59.982±13.883	0.786	0.03
NOS Betweenness Centrality	166.410±25.291	167.660±26.574	0.090	−0.14

Measures are shown across unweighted (binary) networks, with diagnostic group differences in global density, path length, efficiency, and betweenness centrality; weighted networks; at *p_FDR < .05. FA=Fractional Anisotropy. NOS=number of streamlines.

Permutation-based subnetwork analysis

Edge-level analysis identified a differentially connected subnetwork (FA-weighted) in BD, relative to controls (t>1.5, d=.2; p_FWE=.0002) involving 27 structural dysconnections predominantly between and within limbic and basal ganglia nodes, and cerebellar connections via limbic system nodes (Figure 3, Table S9). A differentially connected subnetwork (NOS-weighted) was also seen for BD relative to controls (t>10, d=.5; p_FWE=.005) involving 37 structural dysconnections encompassing connections between default-mode/salience network nodes, and fronto-limbic system and basal ganglia nodes (Figure 3, Table S10). No significant (FA/NOS-weighted) weaker/stronger subnetwork was identified when comparing males to females, or when we tested for a diagnosis-by-sex interaction.

Figure 3. Subnetwork Graph Components Showing Altered FA-weighted and NOS-weighted Connectivity in BD.

The left panel displays a network of altered FA-weighted connectivity in BD compared to controls, with the most significant differences observed between and within the basal ganglia and limbic connections (t > 1.5, p_FWE = .0002, d = .2). Connections with lower FA-weighted connectivity are shown in yellow, while those with higher FA-weighted connectivity are depicted in red. The right panel illustrates altered NOS-weighted connectivity in BD versus controls (t > 10, p_FWE = .005, d = .5), primarily involving the default-mode/salience network regions, fronto-limbic system, and basal ganglia connections. Those connections with lower NOS-weighted connectivity are depicted in blue, while those with higher NOS-connectivity are displayed in green. SupFrontG = Superior Frontal Gyrus; Cau = Caudate; Amy = Amygdala; Hipp = Hippocampus; Thal = Thalamus; FusG = Fusiform Gyrus; VenDC = Ventral diencephalon; Acc = Accumbens; Put = Putamen; Pall = Pallidum; rosAntCingG = Rostral Anterior Cingulate Gyrus; medOrbFronG = Medial Orbitofrontal Gyrus; rosMidFronG = Rostral Middle Frontal Gyrus; supFronG = Superior Frontal Gyrus; tempP = Temporal Pole; Ins = Insula; paraHipp = Parahippocampal Gyrus; PostCingG = Posterior Cingulate Gyrus; LingG = Lingual Gyrus; periCalcG = pericalcarine Cortex; latOccG = Lateral Occipital Gyrus. C=Cortex.

Clinical associations within the Bipolar Disorder Group

At the whole-brain level, longer illness duration (mean±SD: 16±11 years) was associated with lower network density, efficiency, and longer path length (Figure S3a). Later illness onset (21±9 years) was linked to longer path length and poorer efficiency (Figure S3b). A history of psychosis (N=116) was associated with lower density, longer path length, and higher betweenness centrality compared to those without a psychosis diagnosis (N=112) (Figure S3c). All associations remained significant after adjusting for medication use, apart from illness duration effects on centrality. The number of manic or depressive episodes showed no significant associations with global metrics (p_FDR>.05). Statistical details are provided in Table S11.

In subnetwork analyses, longer illness duration was linked to lower FA-weighted connectivity between the right hippocampus and amygdala, and lower NOS-weighted connectivity between the right cerebellum and thalamus (Figure S3d; Table S11). Later illness onset was associated with lower NOS in connections involving the right cerebellum–thalamus, right amygdala–accumbens, right pallidum–insula, amygdala–fusiform, and left amygdala–medial orbitofrontal cortex (Figure S3d). Greater number of manic episodes was associated with higher NOS-weighted connectivity involving the right anterior middle frontal cortex-ventral diencephalon area, while depressive episodes and psychosis history showed no significant subnetwork associations (p_FDR>.05). Results remained significant after adjusting for medication use. No connectivity (whole-brain and subnetwork) differences were observed between diagnostic BD subtypes I (N=201) and II (N=42).

Treatment Associations

In BD, antidepressant use at time of scan was associated with lower network density, longer path length, lower efficiency, and higher betweenness centrality at the whole-brain level, even after adjusting for illness severity and concurrent medications (Figure S4). No significant associations were observed for lithium, anticonvulsants, or antipsychotics on global network measures (p_FDR>.05). At the subnetwork level, anticonvulsant use was associated with higher NOS-weighted connectivity between the right pallidum and both the right ventral diencephalon and superior frontal cortex, and lower connectivity between the right putamen and medial orbitofrontal cortex (Figure S4). Additional patterns of lower NOS-weighted connectivity were observed between the right thalamus and the fusiform, lingual, and lateral occipital cortices (Figure S4). However, after accounting for illness severity, the associations previously observed with anticonvulsant use were no longer significant. No significant FA/NOS-weighted subnetwork associations were found for lithium, antidepressants, or antipsychotics.

Use of serotonin (5HT) reuptake inhibitors (NbN4) was associated with lower density, longer path length, lower efficiency, and higher betweenness centrality at the whole-brain level, even after adjusting for illness course severity measures and other NbN-based treatment categories (Figure S4). 5HT-reuptake inhibitor use was also linked to lower FA-weighted connectivity between the right thalamus and hippocampus (Figure S4), independent of depressive episodes and other medication classes, but not manic episodes. Use of dopamine, serotonin, and noradrenaline receptor antagonists (NbN2) was associated with lower NOS-weighted connectivity between the left and right middle orbitofrontal gyri, even after accounting for illness course severity scores and other NbN categories (Figure S4). Statistical details are provided in Table S11.

Discussion

Our study represents the largest analysis of brain network organization in BD to date, analyzing 450 predominantly euthymic individuals with BD relative to 509 controls across 16 international cohorts. BD was associated with altered whole-brain connectivity, and regional dysconnectivity within fronto-limbic and basal ganglia pathways and networks linking default-mode, salience, and basal ganglia regions. Illness severity measures further modulated these alterations: longer illness duration and later onset were associated with more pronounced global and subnetwork disruptions, a history of psychosis with greater deviations in global network features, and a greater number of manic episodes with higher fronto-limbic connectivity. Pharmacological treatments showed distinct, class-specific associations: antidepressant use, particularly selective serotonin reuptake inhibitors (SSRIs), was linked to lower global network integration (lower density and efficiency), longer paths, higher centrality, and reduced microstructural organization between key limbic regions, while anticonvulsants and antipsychotics were associated with alterations in basal ganglia–mediated emotion regulation and cognitive control circuits, including fewer streamlines between frontal cortices.

Our findings reveal a subtle but reliable pattern of whole-brain dysconnectivity in BD across key measures of network integration and segregation (Figure 2), with the largest group differences emerging for characteristic path length and global efficiency. The magnitude and direction of these effects were comparable to those reported in large-scale meta-analytic work (8), reinforcing the presence of modest yet consistent disruptions in global network topology. Lower global density and efficiency indicate weaker large-scale integration, while longer paths and elevated betweenness centrality suggest that information flow becomes less direct and more reliant on a limited set of hubs. This configuration reflects a compromised topological organization in BD, one in which fewer hubs carry a disproportionate share of network communication. Clinically, reduced integration and hub overreliance may contribute to cognitive and affective disturbances (31) and increase vulnerability to destabilization during stress or mood-state transitions, consistent with evidence linking compromised connectivity to episode onset (32–34). These findings align with prior reports of reduced clustering and efficiency and altered hemispheric or regional connectivity (6,35–38), supporting a convergent pattern of global topological abnormalities across studies. We also observed sex-related differences, with males showing lower clustering and local efficiency than females, consistent with prior work on sex-specific connectivity (39,40) and reports of sex-linked cognitive variation in BD (41), though these differences did not extend to subnetwork connectivity in our sample.

Whole-brain alterations in BD were driven primarily by dysconnectivity within two subnetworks: (1) fronto-limbic and basal ganglia pathways, including posterior cerebellar projections, and (2) a network linking default-mode, salience, fronto-limbic, and basal ganglia regions (Figure 3). The involvement of default-mode and salience networks in BD points to potential dysfunction in self-referential processing and attentional control (42), while across both subnetworks the most pronounced effects involved basal ganglia connections, underscoring their central role in BD-related structural disruption and in emotion regulation, reward processing, and cognitive control. Differences between FA-weighted and streamline-weighted networks suggest complementary microstructural (e.g., axonal density, myelination) and macrostructural (tract-level) alterations (43,44). The co-occurrence of lower and higher connectivity within these subnetworks is consistent with a pattern of both disruption and compensatory reorganization, in which loss of WM microstructure in core affective pathways may be offset by increased reliance on alternative routes to maintain communication. These divergent patterns indicate that BD involves not only structural disruption but also dynamic, potentially compensatory reorganization aimed at preserving communication across affected circuits. Appreciating these network-level adaptations highlights the need to study both structural and dynamic processes and motivates the search for biomarkers that index these compensatory mechanisms, with the potential to guide future circuit-based interventions in BD. Ongoing efforts within ENIGMA are beginning to bridge such MRI-derived patterns with underlying biological gradients, including cell-type–specific gene expression and neurotransmitter profiles (45,46). The anatomical distribution of effects aligns with prior BD findings, particularly involving the anterior limb of the internal capsule, uncinate fasciculus, ventral amygdalo-striatal projections, cingulum, fornix, and corpus callosum; tracts that link basal ganglia nodes with limbic, frontal, and diencephalic regions implicated in affective regulation (6,35,37,47). Consistent with the largest multicenter ENIGMA-BD DTI study (8), which reported widespread FA reductions with strongest effects in the corpus callosum and cingulum, our multisite tractography and network-level approach shows how these regional abnormalities converge into systems-level alterations. This shift from isolated regions to distributed networks is crucial, as converging evidence suggests that psychiatric disorders arise from disturbances in network organization (48,49), with direct implications for developing circuit-level, mechanism-based interventions.

Our findings also highlight the salience network’s role in BD, particularly its integration with subcortical regions involved in interoception and visceromotor control (50). Within this network, the anterior cingulate cortex (ACC) occupies a strategic connector-hub position linking the amygdala, orbitofrontal cortex, olfactory cortex, and temporal regions critical for emotion regulation and homeostasis (51). Dynamic instabilities in interoceptive networks may underlie the maladaptive responses to emotional stimuli observed in BD (15), and consequent aberrant perception of emotional stimuli as increasingly salient (52). The elevated betweenness centrality observed in BD, together with prior evidence of altered connector-hub organization (21,52), suggests increased reliance on these hubs to compensate for lower structural connectivity within core affective and interoceptive pathways. Such hub overreliance may reflect compensatory engagement of connector hubs that support the integration of external stimuli with self-referential processing–a process known to be disrupted in BD and other mood disorders (15,53–55)–and may be key to the regulation of emotional experiences in BD.

Clinical features were significantly associated with topological network organization in BD (Figure S3). Longer illness duration was linked to widespread reductions in network integration across large-scale systems, and older individuals with BD showed greater amygdala–hippocampal connectivity alterations, consistent with cumulative disease burden and evidence that chronic neuroinflammatory processes and glial activation contribute to progressive WM abnormalities in BD, though cross-sectional design limits causal inference (56). Later illness onset was associated with more pronounced global topological alterations, particularly within cerebello-thalamic and fronto-limbic-basal ganglia pathways including reduced connectivity along uncinate fasciculus-related regions. This association warrants cautious interpretation: later onset may reflect differences in treatment timing, illness recognition, or underlying biological and clinical factors, including genetic or experiential reserve, that allow individuals to tolerate illness burden before symptom emergence, though these possibilities cannot be directly tested in our cross-sectional design. Notably, earlier onset is often associated with greater clinical severity, and distinct neurobiological subtypes of BD may contribute to variability in illness onset. A history of psychosis was linked to a more fragmented and inefficient global network, consistent with greater reliance on a limited set of high-betweenness nodes to compensate for broader connectivity deficits (59,61). In contrast, a greater number of manic episodes was associated with higher connectivity between the anterior middle frontal cortex and subcortical regions inferior to the thalamus (hypothalamus, mammillary bodies, subthalamic nuclei, substantia nigra), consistent with mania-induced neural plasticity on fronto-limbic circuitry via dopaminergic and glutamatergic mechanisms (34,57,58,60). The absence of similar effects for depressive episodes suggests distinct mood-state–dependent mechanisms, though our predominantly euthymic sample and cross-sectional design may limit detection of acute or residual mood-related structural changes (62). These results demonstrate that illness-related factors (duration, onset age, psychosis history, and episode frequency) must be modeled to understand BD network pathology, and motivate longitudinal, deeply phenotyped studies to distinguish illness burden, compensatory reorganization, and cumulative manic effects – a distinction with direct relevance for early, mechanism-guided intervention. Richer phenotypic characterization, including symptom clustering, illness trajectory modeling, and item-level severity measures, may further clarify whether observed alterations reflect shared circuit vulnerability or biologically distinct BD subtypes.

Antidepressant use, particularly of selective serotonin reuptake inhibitors (SSRIs), was associated with lower connectivity within limbic circuits involving the thalamus and hippocampus and with reduced global network integration. While SSRIs may help stabilize mood, their mechanism of action could differentially modulate WM connectivity in regions already vulnerable to illness severity in BD. Lower connectivity in specific networks does not necessarily indicate a detrimental effect on brain function. Pharmacological treatments may dampen pathological hyperconnectivity or rebalance circuit dynamics within emotion-regulation networks (65,66). Medication use was not randomly assigned, and although our models accounted for symptom severity and number of depressive episodes, residual confounding and unmeasured aspects of illness chronicity may persist, particularly among individuals more likely to be prescribed SSRIs. Connectivity reductions may therefore reflect adaptive modulation rather than structural impairment and may partly reflect differences in underlying symptom-severity profiles difficult to disentangle without item-level measures, an area of ongoing ENIGMA-BD work. ENIGMA studies of major depressive disorder found cortical and hippocampal differences in medicated patients but no WM microstructural changes (63), suggesting that antidepressant-related WM alterations may be BD-specific illness burden–dependent. Anticonvulsant use produced mixed connectivity changes within emotion-regulation and cognitive-control circuits overlapping with BD social-cognition networks (32) and regions sensitive to anticonvulsant-related cognitive effects (64); these associations did not survive illness severity adjustment, underscoring the challenge of separating medication effects from disease burden and the need for prospective, dose-resolved longitudinal designs. Antipsychotic use was linked to lower connectivity between the medial orbitofrontal cortex and other frontal regions, potentially reflecting therapeutic modulation of limbic–frontal integration or cognitive side effects, with longitudinal dose-resolved data needed to distinguish these. Lithium and other mood stabilizers did not show clear associations with network connectivity in our sample. This absence may reflect lithium’s distinct intracellular signaling mechanisms differing from the receptor- or ion-channel–based actions of SSRIs, antipsychotics, and anticonvulsants, and is consistent with its proposed neuroprotective influence (16,17), suggesting a stabilizing effect on WM architecture (even if not significant here). The relative preservation of network organization in lithium-treated individuals offers a translational distinction between medications that maintain structural connectivity and those associated with reduced integration, suggesting that connectomic markers could help differentiate treatments by their network impact. Whether this reflects a true mechanistic distinction from other psychotropics or methodological limitations requires further study. Overall, our findings support the notion that psychotropic medications influence network organization in circuits aligned with their mechanistic targets, raising important questions about whether these changes are adaptive, maladaptive, or necessary clinical trade-offs; though longitudinal studies remain essential to disentangle medication effects from illness progression.

Several limitations should be considered. While ENIGMA applies standardized processing pipelines and rigorous quality control procedures (18), and linear mixed-effects models with site as a random effect account for multi-site heterogeneity in acquisition parameters and scanner platforms, it is possible that there remains site-related variability. Even so, we note that the heterogeneity is present whether or not the study design models it, and a single site design simply ignores it, by being unable to determine whether the effects are specific to the cohort or if they generalize to other cohorts. Clinical data were standardized across cohorts to accommodate differences in recruitment strategies, diagnostic assessment procedures, and inclusion/exclusion criteria. Medication exposure was assessed cross-sectionally at the time of scanning reflecting naturalistic polypharmacy patterns in BD. The number of individuals receiving lithium monotherapy was relatively small and unevenly distributed across cohorts, limiting statistical power for reliable monotherapy analyses.

Future large-scale, treatment-stratified datasets will be needed to clarify medication-specific network effects. As a result, causal inferences regarding treatment effects on structural connectivity cannot be drawn. Finally, diffusion tractography provides indirect WM pathway reconstruction and is sensitive to crossing fibers, algorithmic choices, and threshold parameters (67,68). Despite these limitations, the large, harmonized sample and standardized analytic framework support the generalizability of observed network signatures across diverse sites.

In the largest BD connectome study to date, we demonstrate the feasibility and value of large-scale, harmonized diffusion MRI connectome analyses, identifying subtle but robust alterations in emotion-regulation and reward-related networks further shaped by illness burden and pharmacological treatment. We report mechanism-of-action-dependent medication effects, particularly antidepressant-related limbic connectivity alterations, underscoring the importance of accounting for treatment exposure and developing more sophisticated approaches to modeling medication effects when interpreting network abnormalities. More broadly, these results illustrate how network-based approaches capture system-level disruptions not apparent in regional analyses and provide a framework for disentangling pharmacological from core neurobiological features of BD. Although cross-sectional design precludes causal inference and observed connectivity differences likely reflect a combination of disease mechanisms, illness progression, treatment exposure, and environmental influences, the consistent involvement of fronto-limbic and reward-related circuits highlights network-level processes relevant to mood regulation and treatment response. Large-scale connectomic studies may identify circuit-level markers to support illness staging, improve treatment stratification, and guide the development of more targeted, circuit-informed therapeutic strategies. Collaborative efforts at this scale will be critical for advancing biomarker discovery and improving the biological understanding of psychiatric disorders.

Supplementary Material

Supplement Description:

Supplement Methods, Figures S1-S4, Tables S1-S11

KEY RESOURCES TABLE.

Resource Type	Specific Reagent or Resource	Source or Reference	Identifiers	Additional Information
Add additional rows as needed for each resource type	Include species and sex when applicable.	Include name of manufacturer, company, repository, individual, or research lab. Include PMID or DOI for references; use “this paper” if new.	Include catalog numbers, stock numbers, database IDs or accession numbers, and/or RRIDs. RRIDs are highly encouraged; search for RRIDs at https://scicrunch.org/resources.	Include any additional information or notes if necessary.
Antibody
Bacterial or Viral Strain
Biological Sample
Cell Line
Chemical Compound or Drug
Commercial Assay Or Kit
Deposited Data; Public Database
Genetic Reagent
Organism/Strain
Peptide, Recombinant Protein
Recombinant DNA
Sequence-Based Reagent
Software; Algorithm
Transfected Construct
Software for Data Processing/Statistics				FreeSurfer v5.3.0; ExploreDTI v4.8.6; NBS v1.2; R v4.2.1