Affect-Regulation Related Emergent Brain Network Properties Differentiate Depressed Bipolar Disorder from Major Depression and Track Risk for Bipolar

Abstract

Background:

Individuals with/at risk for Bipolar Disorder (BD) often present initially for the treatment of depressive symptoms. Unfortunately, pharmacological treatments for Major Depressive Disorder (MDD) can be iatrogenic, precipitating mania which may not have otherwise occurred. Current diagnostic procedures rely solely on self-reported/observable symptoms, and thus alternative data sources, like brain network properties, are needed to supplement current self-report/observation-based indices of risk for mania.

Methods:

Brain connectivity during affect maintenance/regulation was examined in a large (N=249), medication-free sample of currently depressed BD (n=50) and MDD (n=116) patients and healthy controls (n=83). BD risk was categorized in a subset of MDD patients. We used graph theory to identify emergent network properties that differentiated between (i) BD and MDD and (ii) MDD patients at high and low risk for BD.

Results:

BD and MDD differed in the (i) extent to which rostral anterior cingulate was embedded in the local network, (ii) amount of influence hippocampus exerted over global network communication, and (iii) clarity of orbitofrontal cortex communication. MDD patients at high BD risk showed a pattern of local network clustering around right amygdala that was similar to healthy controls, whereas MDD at low risk for BD deviated from this pattern.

Conclusions:

BD and MDD differed in emergent network mechanisms subserving affect regulation, and amygdala properties tracked BD risk in MDD patients. If replicated, present findings may be combined with other markers to assess the presence of BD/BD-risk in individuals presenting with depressive symptoms in order to prevent the use of iatrogenic treatments.

Introduction

Mood disorders are the most common and disabling forms of mental illness [1], affecting 1 in 5 individuals each year [2]. At their core, mood disorders are characterized by disturbances in affect regulation, with euphoria/irritability and depression differentiating Bipolar Disorder (BD) and Major Depressive Disorder (MDD), respectively. Although (hypo)manic episodes discriminate BD, individuals with BD (i) spend significantly more time in major depressive episodes (MDE) than in manic/hypomanic episodes [4,5], (ii) typically present with depressed symptoms in treatment settings, and (iii) are often first misdiagnosed with MDD [6,7]. Misdiagnosis is not simply academic, as common psychotropic treatments for depression can precipitate the onset of mania if administered to individuals with/at risk for BD [7,8]. Given that these individuals may never have gone on to develop mania otherwise, it is imperative to develop additional diagnostic indices that can identify BD risk while individuals are in a depressed state.

Neuroscience is a promising method for addressing the phenotypic overlap between currently depressed BD and MDD, and several elegant neuroimaging studies have identified differences between BD and MDD. Meta-analyses indicate that both BD and MDD display amygdala hyperactivity to affective stimuli [9–11], but that this effect is stronger in BD than MDD [12]. However, the majority of work in this area has not considered the inherent complexity of brain networks, the brain’s chief organizational principle. Understanding the way mood-related pathology is instantiated in connections between brain regions, and how those links are embedded in a larger network context, could reveal differences between BD and MDD that are not apparent using traditional brain connectivity approaches. Furthermore, sophisticated network methods, such as graph theory, provide insight into emergent attributes of networks (i.e., properties only apparent when network complexity is considered) [13].

To date, only two studies have examined currently depressed MDD and BD using graph theory, both of which examined resting-state data from medication-free patients. The first study found more connectivity and local clustering (Clustering Coefficient) in BD (n=13) than MDD (n=40) in a largely prefrontal set of nodes [14]. In the second study, MDD (n=31) evidenced greater local clustering than BD (all BD II, n=32) in a different set of regions (e.g., precuneus) [15]. Although useful, these studies are limited by relatively small samples, reliance on resting-state, and a failure to account for the presence of individuals at high risk for BD within the MDD sample.

To address these gaps, we examined network properties in a large sample of medication-free, depressed young adults (n’s: 50 BD [18 BDI, 32 BDII], 116 MDD) and healthy controls (HC; n=83). We sought to identify graph properties that differentiated (i) BD from MDD, (ii) MDD patients at high- and low-risk for BD, and (iii) patients from HC. Although current self-report-based markers of BD risk are flawed, no other method of categorizing high-risk individuals is currently available without the use of large prospective imaging studies. Thus, BD risk was assessed in a subset of the MDD group (n=80) using self-reported subthreshold mania symptoms to identify additional potential risk biomarkers. Specifically, brain differences underlying risk for mania should vary with self-report indices of risk, because these indices are predictive (albeit imperfectly) of future BD. However, brain differences should index additional risk variance not captured by self-report markers, because such brain mechanisms are likely closer to the causal mechanisms leading to BD. Thus, we chose to conceptually ‘bootstrap’ the identification of brain biomarkers of BD risk by using self-report indices, with the intention that such neuromarkers could be used in combination with other indices to provide more reliable methods for identifying BD risk.

To increase the likelihood of recruiting affect-relevant circuits, participants engaged in a continuous-performance emotion-regulation task that involved viewing negatively-valenced pictures while either maintaining or regulating (via cognitive reappraisal) affective responses [16]. Continuous designs are more akin to real-world regulation, which occurs on a rolling basis [17,18], rather than in discrete segments (i.e., block/event-related design).

We examined four graph properties using a conservative correction procedure for multiple comparisons. Each property indexed complementary aspects of a node’s ability to communicate with the larger network. Here, we briefly introduce the three properties for which we observed findings. Clustering Coefficient measures the extent to which a node is part of a tightly interconnected local network (Figure 1A), and the existence of such sub-networks is necessary for different types of specialized processing within the global brain network [19]. Lower levels of amygdala Clustering, for example, could indicate disruption/inefficiency in the processes supported by the amygdala sub-network (e.g., salience identification). Current-Flow (CF) Betweenness Centrality (Figure 1B) assays the level of influence a node has over information transmitted across the network. For example, decreases in amygdala Betweenness could indicate biologically salient aspects of the environment (e.g., threat) that amygdala is central to identifying [20] have less influence on ongoing processes. Communicability Efficiency (Figure 1C) relates to the quality of the communication a node has with the rest of the network, with higher values reflecting higher quality communication (i.e., less information lost and/or clearer information flow). Thus, lower amygdala Communicability Efficiency could indicate that salience-related information is being degraded when sent from amygdala to the rest of the brain.

Figure 1.

Example Networks Demonstrating Graph Properties

Note: Panel A shows three different networks in which the center (solid black) node has increasing Clustering Coefficient moving right to left. Specifically, in the rightmost network, none of the neighbors of the center node are connected to each other, and thus there is no clustering around the black node. In contrast, all of the neighbors of the black node are connected to each other in the leftmost network, and thus the black node has the maximum level of clustering. Panel B demonstrates Current-Flow Betweenness Centrality. The solid black node has the highest level of Betweenness, because all communication between the horizontally striped nodes and the vertically striped/empty nodes has to flow through the solid black node. In contrast, the empty node has approximately half the Betweenness as the solid node, because it is not as central to the graph structure, despite the two nodes having an equal number of connections. Panel C demonstrates Communicability Efficiency. This property reflects the balance between (i) the number of parallel paths to the other nodes in the network and (ii) the number of self-loops. The greater number of parallel paths, the more that the noise along each path will be cancelled out. The fewer the number of self-loops, the less of the information that is sent out will be wasted. The solid black node has the highest Communicability Efficiency, because it has three parallel paths with which to reach other nodes, but only five self-loops. The striped node has the lowest Communicability Efficiency, because it has only two parallel paths and four self-loops.

Given evidence that amygdala hyperactivity to affective stimuli is evident in both BD and MDD [9–11] (vs. healthy controls), and that this effect is stronger in BD than MDD [12], we hypothesized that patients (vs. HC), BD (vs. MDD), and high-risk (vs. low-risk) MDD would display less adaptive reconfiguration of amygdala network communication in response to the demand for affect regulation. Specifically, given that (i) strengthening the impact of amygdala on network processing would impede affect regulation and (ii) higher levels of amygdala Betweenness, Clustering Coefficient, and Communicability Efficiency would accomplish such strengthening, we predicted that BD, high-risk MDD, and patients overall would evidence relatively higher Betweenness, Clustering, and Communicability during affect regulation.

Materials and Methods

Participants

Participants (N=298) were recruited from Indiana University Hospital/Cleveland Clinic outpatient psychiatry clinics and community advertisements. Procedures were approved by relevant IRBs, and informed consent was obtained. A psychiatrist administered the Mini Neuropsychiatric Interview, and patients satisfied DSM-IV-R criteria for BD or MDD and a current MDE. After exclusion for low data quality, n’s were: total=249, BD=50 (18 BDI, 32 BDII), MDD=116 (35 high-risk, 45 low-risk, 36 unspecified), HC=83. See Supplementary Material for additional information (e.g., exclusion criteria).

Ascertainment of BD Risk in MDD

Three psychiatrists independently reviewed information for 80 MDD patients (data for 36 were collected before review commenced) and classified them as high- (hrMDD) or low-risk (lrMDD) via consensus best-estimate agreement [21]. Based on previous work [22–25], hrMDD (n=35) was conservatively defined as having a past episode of (i) euphoric mood with at least 2 mania symptoms and/or (ii) increased irritability with at least 3 mania symptoms. If full mania symptoms were present, then duration was <4 days [26]. All other reviewed MDD cases were classified as lrMDD (n=45). Family history of BD was also collected, and no lrMDD had such a history (i.e., BD family history conformed to our classification).

Emotion-Regulation Task

Separate runs were collected for affect maintenance and regulation conditions. During both, participants were continuously shown negatively-valenced pictures from the International Affective Picture System [27]. Each picture was shown for 15s (21 pictures/scan), and conditions were matched for picture valence/arousal. Participants were instructed to maintain their emotional response to the pictures during the ‘maintain’ scan and continuously regulate during the ‘regulate’ scan using reappraisal techniques taught during a training session before the scan. Techniques included distancing oneself from what was occurring in the image and imagining it was not real. Following each scan, participants rated the aversiveness of the entire picture set (for that condition). Ratings for 10 participants were not included in relevant analyses due to missing data.

MRI Acquisition/Preprocessing/Computation of Graph Properties

Fig. 2 depicts the computation path. See Supplementary Material for detail.

Figure 2.

Data Path

This figure illustrates the data processing path after preprocessing. This processing stream occurs separately for each condition (maintain, regulate).

Statistical Analyses

Graph properties were entered into a 5,000 permutation-based repeated-measures general linear model (rmGLM) in the Graph Theory GLM toolbox v.46 [28] (repeated factor=maintain vs. regulate). One rmGLM was computed per graph property (per node), with three between-subject categorical predictors of interest modeling the differences between (i) MDD vs. BD, (ii) hrMDD vs. lrMDD, and (iii) all patients (MDD and BD) vs. HC. Thus, we tested whether these moderated task effects (e.g., group X task interaction). See Supplemental Material for covariates. Orthogonal coding was used to reduce the impact of unbalanced sample sizes [29]. Specifically, given that the MDD vs. BD, hrMDD vs. lrMDD, and BD subtype categorical predictors were each nested within one level of another predictor (e.g., hrMDD vs. lrMDD was nested within only the MDD level of the MDD vs. BD predictor), orthogonal coding ensures that there is no collinearity among these predictors.

To decrease the likelihood of Type II errors, we only examined prefrontal and subcortical nodes, given that these regions consistently emerge across relevant meta-analyses [12,30], resulting in 111 nodes total. Given the large number of nodes, we took a two-pronged approach to ensure that the risk for false positives was minimized while simultaneously decreasing the risk of missing important effects. First, given the consistency with which amygdala disturbances have emerged for both BD and MDD individually, along with the comparison of BD and MDD [12,30], we examined amygdala properties without correcting for the other nodes in the network. Second, when examining properties for other nodes, we used the false discovery rate (FDR) [31] procedure to correct for multiple comparisons. Specifically, we corrected across both the number of nodes and the 4 properties examined to account for both sources of multiplicity. Critically, we included amygdala in this correction, ensuring that (for non-amygdala nodes) we corrected across all nodes examined. Thus, we conducted (and FDR corrected across) 111 nodes x 4 properties = 444 analyses total.

To ensure that findings were not driven by characteristics that differed between groups, we repeated all analyses with the addition of eight covariates: illness duration, number of past mood episodes, medication history (i.e., treatment naïve vs. past medication use), gender, YMRS, HAMD, and race (dummy-coded). All findings remained significant, indicating that they were not driven by these variables (see Supplementary Material for further information).

To determine which condition drove significant findings, we examined the group effect within each condition. To gain insights into the impact of that each property on affect, we correlated property values with participant ratings (within each condition). For both the group tests within condition and the correlations of property values with ratings, we corrected for multiple comparisons (i.e., the two conditions) by dividing the critical alpha by 2. For effects that emerged from the MDD vs. BD and hrMDD vs. lrMDD comparisons, we tested whether each group (individually) differed from HC (within each condition), to identify groups exhibiting disturbances from this baseline. FDR correction (across conditions and groups) was used for the comparisons to HC to correct for multiple comparisons.

Results

Only significant analyses are reported. Importantly, all findings below remained significant when confounds were accounted for in analyses.

BD vs. MDD

Patient group (BD vs. MDD) moderated the task effect for right rostral anterior cingulate (rACC) Clustering Coefficient (Table 1). As shown in Figure 3A, clustering was relatively higher for the regulate (vs. maintain) condition in BD, whereas the opposite pattern emerged for MDD. Follow-up analyses revealed that BD showed significantly greater rACC clustering than MDD for the regulate condition, whereas no significant difference was observed for maintain. Additionally, for the regulate condition, rACC clustering was higher in BD than HC, whereas it was lower in MDD than HC (Table 2). Finally, rACC clustering correlated positively (across groups) with valence ratings in the maintain condition.

Table 1.

Group X Task Interactions & Follow-Up Analyses

Major Depressive Disorder (MDD) vs. Bipolar Disorder (BD)
Region	Metric	Uncorrected (Corrected) p-value	ηp2	Group		Correlation w/Rating
				Maintain	Regulate	Maintain	Regulate
Right Rostral Anterior Cingulate	Clustering Coefficient	<.001	.066	.262	BD>MDD	.15	−.03
		(<.030)			<.001	.022	.620
Left Hippocampus	Current-Flow Betweenness Centrality	<.001	.055	.102	MDD>BD	−.10	−.03
		(<.001)			.009	.106	.630
Left Orbitofrontal Cortex	Communicability Efficiency	<.001	.049	MDD>BD	.145	.09	−.03
		(<.001)		.005		.147	.633
High (hrMDD) vs. Low Risk Major Depressive Disorder (lrMDD)
Region	Metric	Uncorrected (Corrected) p-value	ηp2	Group		Correlation w/Rating
				Maintain	Regulate	Maintain	Regulate
Right Amygdala	Clustering Coefficient	.0308	.022	.250	.126	.20	.03
		(n/a)				.001	.653
Depressed Patients (D) vs. Healthy Controls (HC)
Region	Metric	Uncorrected (Corrected) p-value	ηp2	Group		Correlation w/Rating
				Maintain	Regulate	Maintain	Regulate
Left Amygdala	Clustering Coefficient	.032	.020	HC>D	.895	.23	.05
		(n/a)		.007		<.001	.458

Note: Entries in the ‘Uncorrected (Corrected) p-value’ column are the uncorrected (top) and FDR-corrected (bottom) p-values for the overall effect (see Supplementary Material for discussion of why amygdala p-values were not corrected for). Entries in the ‘η_p²’ column are the partial Eta squared for overall effect. Entries in the ‘Group’ column are direction of effects (if significant) and are p-values for the relevant group test within each condition. Entries in the ‘Correlation with Rating’ column are the correlations between affect rating and the graph metrics (top number is the correlation, bottom is the associated p-value). For the ‘Group’ and ‘Correlation with Rating’ tests, the critical alpha was divided by 2 (i.e., number of conditions) to correct for multiple comparisons, and thus the p-value for these tests must be <.025 to be significant.

Abbreviations: MDD=Major Depressive Disorder group; BD=Bipolar Disorder group; HC=healthy control group; hrMDD=Major Depressive Disorder group with high risk for Bipolar Disorder; lrMDD=Major Depressive Disorder group with low risk for Bipolar Disorder; D=Depressed (MDD+BD) patient group; n/a=not applicable.

Figure 3.

Graph Properties Differentiating Patient Groups

Panels A-C reflect the difference between the Bipolar Disorder (BD) and Major Depressive Disorder (MDD) groups in the effect of condition (maintain vs. regulate). Panel D reflects the difference between MDD patients at high- and low-risk for BD (hrMDD & lrMDD, respectively). Panel E reflects the difference between patients (MDD+BD) vs. healthy controls (HC) in the effect of condition. Panel F illustrates the locations of the nodes identified. Means for healthy controls in panels A-D are provided for reference, even though this group was not part of the original effect tested (HC was used in follow-up analyses). All graphs reflect the estimated marginal means (i.e., adjusted for the appropriate covariates).

Abbreviations: CF=current-flow; ACC=anterior cingulate cortex; OFC=orbitofrontal cortex; *=groups are significantly different within task condition; ❖=the mean for that group in that condition is significantly different from the mean for HC in the same condition.

Table 2.

Comparisons to Healthy Controls

Effects Emerging from Comparison of Major Depressive Disorder (MDD) to Bipolar Disorder (BD)
Region	Metric	BD				MDD
		Maintain p		Regulate p		Maintain p		Regulate p
Right Rostral Anterior Cingulate	Clustering Coefficient	.144	(.192)	BD>HC		.902	(.902)	HC>MDD
				.001	(.002)			<.001	(.002)
Left Hippocampus	Current-Flow Betweenness Centrality	.455	(.445)	HC>BD		HC>MDD		.053	(.071)
				.012	(.024)	.008	(.024)
Left Orbitofrontal Cortex	Communicability Efficiency	.029	(.058)	.147	(.196)	MDD>HC		.326	(.326)
						.003	(.012)
Effects Emerging from Comparison of High (hrMDD) to Low Bipolar Risk Major Depressive Disorder (lrMDD)
Region	Metric	hrMDD				lrMDD
		Maintain		Regulate		Maintain		Regulate
Right Amygdala	Clustering Coefficient	.102	(.380)	.586	(.586)	.479	(.586)	.190	(.380)

Note: Entries are the directions of effects (if significant) and the uncorrected and FDR corrected (in parentheses) p-values for the comparisons of the mean graph metric value for each group individually against the healthy control group, within each condition. Note that the healthy control group was not a part of the original effect tested. Thus, these tests were conducted to provide insight into which groups are demonstrating deviations from the typical pattern.

Abbreviations: MDD=Major Depressive Disorder group; BD=Bipolar Disorder group; HC=healthy control group; hrMDD=Major Depressive Disorder group with high risk for Bipolar Disorder; lrMDD=Major Depressive Disorder group with low risk for Bipolar Disorder.

Patient group also moderated the task effect in left hippocampus CF Betweenness Centrality (Table 1). As illustrated in Figure 3B, Betweenness was relatively higher for the maintain (vs. regulate) condition in BD, whereas the opposite pattern emerged for MDD. Follow-up analyses revealed that BD showed significantly lower Betweenness than MDD for the regulate condition, but no significant differences emerged for maintain. Furthermore, BD evidenced lower hippocampus Betweenness than HC for regulate, whereas MDD evidenced lower values than HC for maintain (Table 2).

Lastly, patient group moderated the task effect in left anterior-middle orbitofrontal cortex (OFC) (BA11) Communicability Efficiency (Table 1). As seen in Figure 3C, Communicability Efficiency was relatively higher for the regulate (vs. maintain) condition in BD, whereas MDD evidenced similar levels for both conditions. Follow-up analyses revealed that OFC Communicability Efficiency was significantly lower in BD than MDD for maintain, but no significant differences emerged for regulate. Furthermore, for maintain, MDD evidenced higher OFC Communicability Efficiency values than HC (Table 2).

High- vs. Low-Risk MDD

Risk group (high vs. low-risk MDD) moderated the task effect in right amygdala Clustering Coefficient (Table 1). As seen in Figure 3D, clustering was relatively higher for the maintain (vs. regulate) condition in BD, whereas MDD evidenced similar levels for both conditions. Follow-up analyses did not reveal any significant group differences when examined within each condition. Furthermore, right amygdala Clustering Coefficient positively correlated with valence ratings for the maintain condition.

All Depressed Patients vs. Healthy Controls

Patients differed from HC in left amygdala Clustering Coefficient (Table 1). As seen in Figure 3E, clustering was relatively higher for the maintain (vs. regulate) condition in HC, whereas patients evidenced similar levels for both conditions. Furthermore, left amygdala clustering positively correlated with valence ratings for the maintain condition.

Discussion

This study leveraged new methods for indexing emergent properties of brain networks to identify biomarkers that differentiate BD from MDD in a large sample of currently depressed, medication-free patients. Specifically, we applied graph theory methods to fMRI data collected during a continuous emotion-regulation task and identified several potential biomarkers. For example, Clustering Coefficient in right rACC differentiated BD from MDD and tracked patients’ self-report ratings of emotional aversiveness during the task. Furthermore, right amygdala Clustering Coefficient differentiated between MDD patients at high- and low-risk for developing BD and also tracked self-reported aversiveness. These findings demonstrate the promise of using emergent brain network properties to identify novel metrics and improve early identification of depressed patients at high risk for developing BD.

Differentiating BD and MDD

Network clustering (Clustering Coefficient) around right rACC differentiated BD from MDD in the regulate condition (Figure 3A). Moreover, BD exhibited higher levels of rACC network embeddedness than HC when regulating affect, whereas MDD levels were lower than HC. Thus, the local network surrounding right rACC during emotion regulation appears to be most interconnected for BD, followed by HC, then MDD. Because tightly interconnected local networks are crucial for engaging in specialized processing [19], this finding suggests that the processes supported by rACC differ by group during emotion regulation. rACC is crucial for affect-related top-down control (e.g., of amygdala, [32]), suggesting that currently depressed BD patients over-engage top-down control of affect, whereas such control is weaker in MDD. This interpretation is supported by evidence that rACC shows decreased resting-state connectivity with amygdala in unipolar [33], but not bipolar [34], depression.

BD and MDD also differed in the level of influence left hippocampus had over information transmitted in the network (CF Betweenness Centrality; Figure 3C). Relative to HC, hippocampus had less influence (i) in MDD during maintenance of affect and (ii) in BD during regulation of affect. However, only MDD showed a flip in the condition effect relative to HC, suggesting that this disturbance is more pronounced in MDD. Given the central role of hippocampus in contextual binding [35], weaker hippocampal influence on network communication in MDD may lead to weaker context-based modulation of ongoing processing. This is supported by the negative correlation between affect ratings and Betweenness during the maintain condition. If true, processing of negative stimuli may not be limited to relevant contexts in MDD individuals, potentially contributing to depressive rumination (i.e., repetitive negative thoughts, even when not contextually relevant) [36].

Finally, the quality (more information transmitted/less information wasted) of left anterior-middle OFC communication (Communicability Efficiency; Figure 3C) was lower during the maintain condition for BD and HC, whereas levels were higher in both conditions for MDD (i.e., similar to regulate for BD/HC). This suggests that BD and HC are evidencing lower clarity of OFC communication during unregulated affective processing, whereas higher clarity communication was evident for MDD in both conditions. Given OFC’s role in maintaining stimulus value in the current context [37], higher quality OFC communication may reflect engagement in the trained reappraisal strategies, which involved reinterpreting stimuli by changing the context (e.g., telling oneself that individuals in a violent photo are actors and not in danger). If so, our finding that MDD shows higher quality OFC communication during both conditions may reflect the fact that these individuals are continuously regulating affect, consistent with past findings of lower affect reactivity in MDD [38]. Conversely, the fact that MDD is not differentiating between conditions could contribute to impairment.

Biomarkers of BD Risk in MDD

Right amygdala network embeddedness (Clustering Coefficient) differentiated between MDD patients at high (hrMDD) and low (lrMDD) BD risk (Figure 3D). Contrary to hypothesis, hrMDD evidenced a similar pattern as HC, namely greater Clustering during maintain, relative to regulate. In contrast, lrMDD showed no difference between conditions, suggesting that ‘true’ MDD patients fail to modulate amygdala Clustering appropriately based on context. Given that the lrMDD means for both conditions lie about midway between the means for both hrMDD and HC, it is not the case that the disturbances observed in lrMDD are limited to one condition.

Interestingly, amygdala Clustering during affect maintenance was positively related to self-reported aversiveness of task stimuli. Thus, such clustering may support processing of the aversive aspects of stimuli, consistent with the central role for amygdala in determining motivational salience [20]. As mentioned above, network clustering is crucial for supporting specialized processing within sub-circuits. For example, higher Clustering around amygdala may reflect greater elaboration of amygdalar processing, because the regions that receive information from amygdala also transmit to each other, and thus amygdalar information may be passed back and forth to a greater extent. Therefore, higher Clustering during affect maintenance may drive increases in the detection of salient (i.e., aversive) stimulus features, consistent with the positive correlation between Clustering and aversiveness ratings. When considered in this context, this suggests that individuals with ‘true’ MDD (lrMDD) identify aversive salience similarly across contexts, which could impede emotion regulation. In contrast, those at high risk of BD (hrMDD), may identify salience similar to healthy individuals, at least when depressed.

Depression Across Disorders

Across mood disorders, depressed patients exhibited less differentiation (vs. HC) in left amygdala Clustering Coefficient between task conditions (Figure 3E). Specifically, HC evidenced increased amygdala Clustering during the maintain condition (vs. regulate), but depressed patients did not exhibit this increase. This may appear counterintuitive at first, as depression is typically conceptualized as higher in negative affect. However, depressed individuals often exhibit lower state responsivity to negatively-valenced stimuli, despite consistent negative mood [38]. Thus, decreased amygdala Clustering may reflect one mechanism by which reactivity to aversive stimuli is attenuated in depression.

Implications

One interpretation of the present findings is that they lead to/maintain (hypo)mania, as this is what differentiates BD from MDD. At the same time, because all patients were currently depressed, present findings may instead reflect differences in the way depressive symptoms are instantiated in BD vs. MDD. Specifically, although they may experience the same surface symptomology, the mechanisms which cause/support such pathology may differ between disorders. Importantly, present findings are not due to differences in the severity of current depression or mania, as all findings remained significant after controlling for HAMD and YMRS.

One clue that may assist in differentiating between these explanations is comparisons between property levels/patterns of the patient groups to HC. For example, BD and HC showed a similar pattern (regulate>maintain) of OFC Communicability Efficiency and there were no significant differences in property levels, whereas MDD showed a different pattern (regulate=maintain) and higher property levels during maintain than both HC and BD. Therefore, this neuromarker seems to be driven by disturbances in MDD, rather than BD, and thus may be a mechanism supporting depression uniquely in MDD. Similarly, right amygdala Clustering Coefficient in MDD patients at high risk for BD showed a similar pattern (maintain>regulate) to HC, whereas MDD at low BD risk showed a different pattern (maintain=regulate). Thus, this may be a mechanism unique to ‘true’ depression in MDD, and the absence of this pattern in current depression may indicate higher BD risk. No finding was particularly indicative of the first explanation mentioned above (i.e., that the neuromarker supports mania in some way). For example, the disturbances were evident in both groups’ levels of right rostral ACC Clustering Coefficient, and thus this neuromarker is likely to support pathology in both. Thus, future research integrating currently euthymic BD/MDD and/or (hypo)manic BD is needed to accurately parse these potential roles for the observed findings.

The present findings may serve as useful neuromarkers for differentiating currently depressed individuals with/at high vs. low BD risk, even without understanding how these findings support BD and/or MDD. For example, even if the observed differences are indicative of MDD ‘proneness’ rather than BD, the absence of such markers in currently depressed individuals could still serve as a useful indicator of BD risk. The identified neuromarkers could be used in combination with other risk markers (e.g., subthreshold mania symptoms) to guide treatment decisions for those presenting with current depressive symptoms. Future research (e.g., using machine learning) is needed to determine the optimal contribution of each risk marker for making this determination.

Strengths, Limitations, and Future Directions

This study has a number of strengths, including a large sample size (n=249), medication-free patients (quite uncommon in BD research), and a novel continuous-performance emotion-regulation paradigm. We also employed a cutting-edge analysis framework, investigating graph properties that were appropriate for use in brain networks and accounting analytically for the influence of lower-level network characteristics. Furthermore, this is the first study to examine BD risk in MDD using graph properties.

Several limitations must be considered. First, we collapsed across BD subtypes (although subtype was covaried to account for this variance). Future studies, with larger samples, should test for subtype differences. Second, although participants were at least 2 weeks medication-free, many had previously taken medication, which can have normalizing effects on network properties [39]. Third, this study was cross-sectional, and thus longitudinal work will be key to understanding the predictive validity of findings. Fourth, the order of the task conditions was not counterbalanced, which could have had several effects (e.g., fatigue, carryover effects). Finally, although average age of the MDD patients (26 yrs.) was only slightly above the average age at onset of BD (23 years [40]), it is possible that sampling at this age reduced the likelihood of obtaining individuals who will actually go on to develop BD (e.g., these individuals would have already developed BD by this age, if they were going to). In fact, these individuals might be somewhat ‘resilient’ against developing BD.

Overall, this study provides some of the first insights into how depression is instantiated in emergent brain network properties and how such properties can differentiate BD from MDD, along with those at risk for BD. We found that BD and MDD evidenced differences in brain network mechanisms subserving depression-related disturbances in affect regulation, including in regions critical for both affect production and regulation (e.g., amygdala, OFC). Moreover, we identified neuromarkers that differentiated high and low risk for BD in currently depressed individuals who have not yet experienced (hypo)mania. Detecting risk markers for BD in depressed individuals is crucial given the lack of diagnostic tools for differentiating these patients and the life-changing consequences incorrect diagnoses can have for precipitating BD in currently unaffected individuals. Given that MDD appeared to differ from the typical pattern (i.e., in health controls) to a greater extent than BD in the majority of the observed findings, it is possible that these findings are provide insight into mechanisms specific to depression in MDD. Thus, future research is needed to identify the mechanisms that lead to BD predisposition, in particular.

Table 3.

Summary of Main Findings

Condition	Right rACC Clustering Coefficient	Left Hippocampus CF Betweenness Centrality	Left OFC Communicability Efficiency	Right Amygdala Clustering Coefficient	Left Amygdala Clustering Coefficient
Maintain	ns	ns	MDD > BD	Effect for hrMDD vs. lrMDD, but groups did not differ within condition	HC > Patients
Regulate	BD > MDD	MDD > BD	ns		ns

Note: BD=Bipolar Disorder; CF=current flow; MDD=Major Depressive Disorder; hrMDD=MDD at high risk for BD; lrMDD=MDD at low risk for BD; ns=non-significant difference; OFC=orbitofrontal cortex; rACC=rostral anterior cingulate cortex.