Multiscale neural signatures of major depressive, anxiety, and stress-related disorders

In order to compare neuroimaging signatures of MDD, anxiety, and stress-related disorders, we first calculated the case–control brain maps. Compared to healthy controls, cortical thickness was reduced across many prefrontal, parietal, and temporal regions in the MDD- and MDD + ANX group (Fig. 1E and SI Appendix, Table 3). These differences were also captured in the comparison of all cases vs. controls (Fig. 1A) as the case–control map for all cases was very similar to the case–control map for MDD-, ANX-, and MDD + ANX (Fig. 1 C and E). In contrast, only select temporal and parietal regions showed a significant decrease in cortical thickness in the ANX- group including middle temporal and supramarginal regions. Significantly reduced cortical thickness was also found in the parahippocampal and ventral medial visual regions of the STR- group when compared to controls.

In individual comparisons of each diagnostic group to healthy controls, we identified widespread differences in resting state connectivity (Fig. 2D). The largest number of significant differences was found between the control and MDD- or MDD + ANX groups, whereas smaller differences were found between the control and ANX- or STR- groups. Specifically, independent component (IC)–12 (presupplementary and supplementary motor areas) showed decreased connectivity with other motor areas (IC-17, superior temporal gyrus) and increased connectivity with the striatal IC-18 across MDD-, MDD + ANX, and ANX- groups but not STR-. A comparison of all cases to controls (Fig. 2A) showed that the pattern of differences across all cases was very similar to the differences observed specifically in MDD-, ANX-, and MDD + ANX (Fig. 2B).

Case–control differences in cognitive function are summarized in Table 2 and SI Appendix. Compared to the control group, the four diagnostic groups showed significant impairments in all domains of cognitive function, except for visuospatial processing (trail-making test [TMT]), which was not significantly impaired in STR-, and PAL, which was not significantly impaired in MDD-, MDD + ANX, or STR-. Notably, we found impaired fluid intelligence and digit–symbol substitution performance in all case groups, suggesting a transdiagnostic pattern of impairments among these aspects of cognitive function.

After constructing diagnosis-specific case–control neural signatures, we evaluated the pairwise similarities in their spatial distributions and component effect sizes. A large degree of overlap was found between MDD-, ANX-, and MDD + ANX groups in both functional connectivity and cortical thickness case–control associations (Fig. 2 B and C). On the other hand, the STR- group showed a different spatial pattern of associations that did not resemble the other disorders (P_PERM > 0.01).

The similarities of neural signatures of the disorders represent the shared influences of both heritable and environmental factors. To determine the degree to which these similarities were moderated by interindividual differences in heritable risk for each of these comorbid disorders, we performed the same analysis again but this time including PRS for all three diagnoses as covariates. After controlling for these genetic risk profiles, correlations between diagnostic neural signatures for the MDD-, ANX-, MDD + ANX, and STR- groups were slightly reduced for both functional and structural measures (Figs. 1D and 2C). The correlations between STR- and all other diagnostic groups were affected to the greatest degree compared to other groups.

A comparison of active MDD with the other disorder groups revealed a highly similar pattern of case–control differences to the analysis of the full sample (SI Appendix). Therefore, restricting the sample to active MDD did not alter the disorder similarity matrices. Focusing on nonmedicated participants also did not substantially alter the results, suggesting that our main findings are likely driven by the presence of lifetime diagnosis rather than medication.

Given the substantial, albeit variable, overlap in functional neural signatures between the diagnostic groups, we aimed to uncover those circuits specifically related to worse executive function and verbal memory across all disorders (n = 3,216) after regressing out age, sex, site, and motion confounds. Partial least squares (PLS) regression identified three latent variables that explained 3.1%, 1.5%, and 1.3% of variance in the four cognitive function tests (TMT, fluid intelligence [G_F], PAL, and digit–symbol substitution task [DSST]). Permutation testing showed that these latent variables together explained a significant amount of variance in cognitive performance (P_PERM < 0.001; Fig. 3A).

The first latent variable (PLS1) representing functional connectivity, which was optimally associated with cognitive performance, captured the most variance in the outcome variables (Fig. 3A) and was associated with worse performance on all four cognitive tests. We found 11 connectivities with normalized PLS1 weights of Z > 3 (significant and positive) and 8 connectivities with normalized PLS1 weights with Z < −3 (significant and negative) (Fig. 3 B and C). Functional correlates of worse cognitive performance included connectivities of the frontoparietal, default mode, and attention networks. Increased DMN–FPN connectivities were also associated with worse cognitive function (e.g., IC-6 with IC-14 or IC-20 with IC-6). Decreased within-DMN (IC-20 with IC-14 or IC-7 with IC-9) and within-FPN (IC-5 with IC-21 or IC-5 with IC-16) connectivities were associated with worse cognitive function.

The second latent variable (PLS2) captured some variance in fluid intelligence and PAL (SI Appendix, Fig. S10). We found four connectivities with significant positive loadings on PLS2 (Z > 3) and one weight with significant negative loading (Z < −3). Higher connectivity of IC-21 (FPN/DMN) with IC-5 (FPN) and IC-1 (DMN) and higher connectivity of the striatal component (IC-18) with the superior temporal gyrus (IC-17) predicted better fluid intelligence and associative learning. Lower connectivity of IC-21 (FPN/DMN) with IC-6 (FPN) predicted lower fluid intelligence and associative learning scores. Although the third latent variable (PLS3) was significant at permutation testing, there were no PLS3 weights |Z| > 3.

To contrast transdiagnostic signatures with diagnosis-specific signatures, we repeated the PLS regression in each case group separately. We found significant brain–cognition relationships in MDD- and in ANX- but not in MDD + ANX or STR- (Fig. 4 A and B and SI Appendix), suggesting that former two groups drive the associations across all cases. We found a distinct pattern of brain–cognition relationships in MDD- and ANX-. While components from the FPN and DMN played a major role in brain–cognition relationships in each case group, different connectivities among different components of these networks were implicated in MDD- and ANX-. Many components that were significant (|Z| > 3) in MDD- and ANX- separately were also significant in the PLS regression across all cases.

In order to directly compare brain–cognition associations in MDD- and ANX- post hoc, we used univariate Pearson’s correlations between functional connectivities and the first principal component capturing the largest amount of variance in Y (46%). We found several connectivities that were significantly different between MDD- and ANX-, each of which was significantly associated with cognitive function in one group but not in the other group (Fig. 4 C and D).

We repeated the PLS analysis in the control group. Briefly, the PLS model explained 3.6% of variance in the cognitive outcomes (P_PERM < 0.001; SI Appendix). PLS1 was also linked to worse general cognition in healthy controls. PLS1 weights included ICs from the DMN, FPN, and dorsal/ventral attention [DA/VA] network (SI Appendix, Figs. S11 and S12). PLS2 was also linked to worse cognition, similarly to PLS2 in the case group PLS.

The MDD- and MDD + ANX groups showed cortical thickness differences from the control group that were consistent with previous meta-analyses of brain structure in MDD (34, 35), with notable thinning of the parietal and prefrontal cortex (insula, anterior cingulate, inferior frontal, superior frontal and middle frontal gyri, and anterior temporal lobe). Comorbid MDD and anxiety are typically associated with greater symptom severity and more limited level of daily functioning (39, 40). Consistent with the expectation of greater disorder severity in comorbid MDD and anxiety disorders, we found additional thinning in the medial temporal regions (parahippocampal area), suggesting a more profound impact of the comorbid disorders on brain structure. The ANX- group showed localized patterns of cortical thinning in the middle temporal gyri, which is in line with previous evidence of reduced volumes of these regions in anxiety disorders without comorbid MDD (11). Unlike studies of active anxiety disorders, we did not find significant differences in the anterior cingulate or insular cortices in our ANX- group. This may be explained by the possible time gap between diagnosis and imaging data. In the UKB, lifetime diagnosis is recorded, and many participants may not be experiencing clinical symptoms at the time of scanning. One possibility is that brain volumes may have recovered since the time of formal psychiatric diagnosis. Finally, the STR- group showed few significant differences from controls, characterized primarily by localized cortical thinning in the parahippocampal area. Parahippocampal cortex is critical to memory formation and retrieval, and medial temporal areas have been emphasized in studies of PTSD (41, 42). Further, the hippocampus has been shown to regulate stress response by influencing the hypothalamic–pituitary–adrenal function (37). Therefore, our findings in the STR- group are consistent with existing theories of stress regulation in the medial temporal lobe.

Significant resting-state connectivity deviations in people with MDD and MDD with comorbid anxiety were similar and encompassed a variety of regions that were part of the default mode and frontoparietal but also subcortical (striatal), cerebellar, and visual networks. Deviations in connectivity within the motor network and between motor and striatal networks were common across the MDD, anxiety disorders, and MDD comorbid with anxiety groups. Our findings are consistent with previously reported alterations in frontoparietal and default mode connectivity in MDD (31) and in anxiety disorders (32, 33). Connectivity between striatal and cortical regions is known to be impaired in MDD (43) and was also altered in both our MDD groups. Importantly, functional connectivity was significantly different between the anxiety disorders and the control group but not in stress-related disorders, reinforcing the differences in neural signatures of this group. Although the effect sizes of case–control differences were relatively small (0.05 < d < 0.15), they were slightly larger than effects previously reported in large-scale studies of psychiatric disorders (44, 45). It has been shown that small effect sizes in large-scale samples examining brain–behavior relationships tend to be reproducible, compared to larger effects in smaller n studies (46). While it is possible that studying more ill patients might lead to larger effect sizes or different findings, our analyses focused on actively ill patients were similar to our overall results. Effect sizes found in the UKB likely reflect the heterogeneity of this large-scale community-recruited sample and support the investigation of neurobiological dimensions underlying this variability consistent with the research domain criteria approach (16).

Substantial evidence shows that scanner type and site are two of the most important confounds in multicenter studies (47–50). Correcting for such confounds can be challenging (47) and can decrease the effect of diagnosis. Site effects also present a challenge in the UKB data (51). Approaches used by UKB and in our analyses may have mitigated site and scanner effects. First, UKB uses identical scanner hardware and software in all three sites, although subtle differences, e.g., in the same coil equipment cannot be ruled out (51). Second, IC analysis-based denoising of resting-state fMRI data used in UKB diminishes site scan differences (52), potentially improving the power of our fMRI analyses. Third, we controlled for linear effects of site in all analyses and were able to detect moderate but robust effects.

The strong similarity of the neural signatures of the MDD and anxiety disorders groups supports previous studies of shared neural signatures of mental disorders (12). Controlling for polygenic risk somewhat reduced the similarity between MDD, anxiety disorders, and stress-related disorders. High genetic correlation between MDD and anxiety disorders (5, 53) could instead account for phenotypic and clinical similarities between the disorders, and not all genetic similarity may be captured by PRSs. By contrast, stress-related disorders show less genetic similarity to MDD or anxiety disorders (5). Further, the STR- group showed very few differences from healthy controls, suggesting lack of effects rather than a different pattern of neural changes relative to controls is distinguishing STR- from MDD- and ANX-.

Several neurocognitive processes may underlie the overlap in neural signatures of MDD and anxiety disorders, including executive functioning (19). A shared neural mechanism could be found in prefrontal regulation of the default mode and limbic circuits responsible for mood and emotional processing. Stress-related disorders, on the other hand, showed differences in the parahippocampal structure and default mode connectivity that may be linked to maladaptive stress response and memory formation.

We identified impairments in different aspects of cognitive performance, including executive function in the four diagnostic groups, consistent with previous literature suggesting that executive dysfunction may be a transdiagnostic dimension of cognitive impairment (16, 54, 55). Executive function impairments are not unique to mental disorders as physical problems, acute stress, or social isolation can also impact executive performance (56).

While much is known about neural correlates of executive function (57–61), lack of replication and small samples call for large-scale data (>2,000 individuals) to uncover reproducible brain-wide associations with cognition (46). In our large UKB sample, we found a specific functional connectivity profile between nodes of the frontoparietal, default mode, and salience/attention networks and the cerebellum as underlying worse processing speed, spatial attention, fluid intelligence, and PAL. We showed that reduced within-network and increased between-network connectivity of the frontoparietal and default mode networks predicted worse executive function and related cognitive impairments in MDD, anxiety, and stress-related disorders. A similar pattern of brain–cognition relationships was found in the healthy control sample. These findings are consistent with previous multimodal metaanalyses showing that neural correlates of executive function in clinical populations encompass frontal, parietal, and cerebellar regions (60) and previous large-scale analysis relating frontoparietal and default mode connectivity to worse verbal-numerical reasoning (61) in UKB. Successful executive functioning relies on dynamic switching between frontoparietal and default mode networks initiated by cinguloopercular attentional networks (59, 61). Self-referential processing in the default mode network is balanced with goal-oriented solving of complex problems (executive function) and influenced by salient events that can trigger the switching by attention networks.

Executive function itself is multidimensional, with processing speed, cognitive flexibility, fluid intelligence, working memory, planning, attention, and inhibitory control serving as interrelated subdimensions (56). The second latent variable identified in our PLS analysis (PLS2) uncovered functional correlates of poorer performance in fluid intelligence and associative learning. In addition to connectivity pairs already identified as related to general cognitive function by the first PLS latent variable, PLS2 showed that higher connectivity of striatal areas with superior temporal gyri predicted better fluid intelligence and associative learning specifically. Identifying robust neural correlates of cognitive function using resting-state connectivity has been challenging due to the limitations of the sample sizes (46) and the interindividual variability in brain function and its relation to cognitive performance (62, 63). Here we identify robust brain–cognition relationships that are replicable in a healthy population, furnishing evidence for existing theories of how functional organization of the brain is associated with cognitive dysfunction, including executive impairments.

We did not find that brain–cognition relationships identified across all diagnostic groups were necessarily present in each of the groups separately. Only MDD- and ANX- showed significant associations between cognitive performance on all four cognitive tests and functional connectivity captured by the first latent PLS variable. A comparison of the univariate correlations showed that some of the brain–cognition associations were distinct between these two groups. Higher striato–default mode connectivity and lower within-default mode connectivity was associated with worse cognitive function in ANX- but not in MDD-. On the other hand, higher between-network connectivity (e.g., between attention and frontoparietal components) was associated with worse cognitive function in MDD- but not in ANX-. These findings suggest that brain–cognition relationships may not be uniform across MDD and anxiety disorders. Instead, different aspects of frontoparietal and default mode interactions may underlie cognitive dysfunction in MDD and anxiety disorders.

The cross-sectional nature of the data does not allow us to disambiguate whether the neural signatures are the consequence of or a marker of vulnerability to the respective disorder. Although diagnostic groups were defined by lifetime diagnosis, we found that approximately one-third of participants in our MDD or MDD with comorbid anxiety groups had some current depressive symptoms based on the two-item patient health questionnaire (PHQ-2) (38) at the time of cognitive and MRI assessment. We show that our results did not change when restricting the sample to active MDD. PHQ-2 with a threshold of 2 has a high sensitivity but moderate specificity of 0.67 (38) as a screening tool for depression, and many individuals in the active MDD group showed only mild symptoms. Given that the PHQ-2 was generally designed as a screening tool maximizing sensitivity at the expense of specificity, we are less able to draw inference about active illness and conclude that our findings appear to be driven by lifetime diagnosis. Restricting the sample to unmedicated participants also produced consistent findings to the full sample. However, chronicity of the psychiatric disorders may differ given that the diagnosis of MDD- was reported at a younger age compared to ANX- or STR-. Further, diagnoses were ascertained using electronic health records that may show some heterogeneity (64).

While our analysis showed a similar signature of brain structure and function deviations across MDD and anxiety disorders, there are limitations to the degree of overlap between the disorders. No significant deviations in cortical thickness were seen in anxiety disorders, while the effects on cortical thickness seen in MDD and in the comorbid group were much larger and reached statistical significance. Interestingly, neuroimaging correlations accounted for less than 50% of variance in the disorder brain maps, suggesting that there are disorder-specific abnormalities in brain function and structure (65).

In conclusion, we found a high degree of similarity in the neural signatures of MDD and anxiety disorders (alone and in comorbidity) that was distinct from stress-related disorders. Our findings are consistent with the diagnostic categorization of MDD and anxiety disorders as internalizing disorders (7). Stress-related disorders showed a very similar profile of executive dysfunction to MDD and anxiety disorders, yet their neural signatures showed less similarity, especially in the domain of cortical thickness. While the comorbidity across disorders is viewed as a therapeutic challenge, the identified neurobiological substrate of connectivity within and between default mode and frontoparietal networks subserves cognitive dysfunction and could provide a promising target for specific interventions.