1 INTRODUCTION

Large-scale, multisite brain imaging datasets are becoming more common through initiatives such as ENIGMA (McWhinney et al., 2023), ADNI (Cruciani et al., 2024), ABCD (Dahl et al., 2024), the human connectome project (Cohen et al., 2023), and others. Large datasets allow us to apply multivariate techniques of analyses, which model interplay between regions (Woo et al., 2017), but require larger, more ecologically valid samples to provide more replicable results (Marek et al., 2022). These techniques better fit the anatomy of complex neuropsychiatric disorders which are characterized not by alterations in a single region, but rather by variations across distributed brain networks (Hibar et al., 2018; Segal et al., 2023). However, there is little methodological clarity on which of the many available methods of multivariate data analyses are best suited to the task of relating brain structure to system-level variables. While development of new methods is one key aspect of the field, uncovering benefits and best-use scenarios for established methods is equally as important.

Analyzing brain imaging changes in BD is a suitable way to test multivariate techniques. Individuals with BD markedly vary in their clinical presentations and impact of the illness on their functioning. This clinical heterogeneity may reflect neurobiological heterogeneity, which can be studied by brain imaging. It is increasingly clear that brain alterations in severe mental illnesses (SMI) are multifactorial. Aside from the diagnosis, they also reflect the effects of additional clinical factors, including medications (Hajek et al., 2012; McWhinney, Abé, et al., 2022; Van Gestel et al., 2019), and comorbid psychiatric or physical conditions, such as obesity (McWhinney, Abé, et al., 2021; McWhinney, Brosch, et al., 2022; McWhinney, Kolenic, et al., 2021) and diabetes (Hajek et al., 2014, 2016). Understanding the brain changes in SMI and translating these findings into clinical settings requires sensitive and replicable methods that link patterns of brain alterations to system-level variables.

Broadly speaking, some methods, such as clustering, categorize participants into groups based on their brain structure, while others, such as principal component analyses, represent brain imaging data as a linear combination of features. While clustering has become a popular method for multivariate analyses of neuroimaging data (McWhinney, Abé, et al., 2022), we do not expect groups of individuals to fall neatly into distinct clusters (e.g. healthy vs. unhealthy). Also, external variables may not exhibit a binary effect on the brain, but rather a nuanced, continuous one. Indeed our previous study using clustering exemplified these issues. We found that there were no strictly separate clusters in brain imaging data and that the boundaries between BD and controls were not clear, i.e. many controls fell into the cluster together with BD individuals, while some BD individuals clustered with controls (McWhinney, Abé, et al., 2022). The cluster assignment of individuals in part depended on continuous variables including age and BMI and effectively resulted in categorizing of these continuous variables, which is not optimal.

Neuroimaging data are often strongly correlated naturally (i.e. brain networks) and due to preprocessing (i.e. coregistration, smoothing, etc.), which is a good reason for linear projection methods. Instead of categorization, such methods quantify degrees of variation and may be better suited to identifying sources of heterogeneity in brain imaging data, as many of these sources may in fact themselves be on a continuum. While machine learning (ML) techniques can overcome these challenges, such techniques require large training sets and out-of-sample validation, and results can be difficult to interpret and translate into practice. Principal component analysis (PCA) represents a potentially optimal middle ground between these approaches, as it can perform well using modest sample sizes while reliably reducing dimensionality across many variables (i.e. brain regions) and deriving robust low-dimensional data representations (Comrey & Lee, 2013). By identifying covariance across individuals in numerous regions simultaneously, PCA can identify patterns of distributed brain network changes that can subsequently be linked with clinical correlates, while maintaining interpretability (Behdinan et al., 2015; Maralakunte et al., 2023; Rehák Bučková et al., 2023; Yeh et al., 2010).

Our main goal here is to compare whether methods which represent brain imaging data as a linear combination of features are better in capturing associations with clinical variable than methods which categorize brain imaging data into clusters. To do that we selected the most established and representative examples of each approach, i.e. PCA vs K-means clustering. Specifically, we used PCA to identify patterns of covariance across regions of interest and related them to clinical and demographic variables. We then compared performance of this approach to our prior clustering study on identical sample. We expected that compared to clustering, patterns of covariance in brain imaging data, as identified by PCA, would show stronger associations with clinical and demographic variables.

2 MATERIALS AND METHODS

3 RESULTS

3.2 Principal component analysis

In the primary sample, the scores for each of the first PCs accounted for 42.7% of variance in cortical thickness, and 46.2% in surface area (see Figure 1). With exception of the second component for cortical thickness (11.5%, see Figure S1), all other components in both measures explained <5% of variance each. The first principal component scores were associated with higher cortical thickness and surface area in all studied regions, with some regional variations in the strength of association (see Figure 2). Consequently, if a clinical variable was associated with lower first component score, then it would be associated with lower cortical thickness across regions in a pattern reflecting the component's loadings across the regions.

For cortical thickness, the first component's score was negatively associated with diagnosis of BD, BMI, and age, indicating that BD, BMI, and age were independently associated with a diffuse pattern of thinner cortex, see Table 2. Among participants with BD, lithium and antipsychotic medications showed opposing associations with the first principal component, such that antipsychotics were negatively while lithium was positively associated with cortical thickness, see Table 2. The second component for cortical thickness was significantly associated with BMI, age, and sex, see Table S7.

TABLE 2. Associations between demographic, clinical, and treatment characteristics and the first principal component of cortical thickness and surface area.

		All participants		Participants with BD
		Estimate (SE)	Significance	Estimate (SE)	Significance
Cortical thickness	Group (BD)	−1.38 (0.17)	F(1,2421) = 67.80, p = .000*	n/a	n/a
	BMI	−0.39 (0.13)	F(1,2419) = 8.79, p = .003*	0.25 (0.29)	F(1,436) = 0.75, p = .387
	Sex (F)	0.16 (0.14)	F(1,2418) = 1.30, p = .254	−0.01 (0.34)	F(1,435) = 0.00, p = .970
	Age	−3.10 (0.10)	F(1,2422) = 1003.34, p = .000*	−3.37 (0.28)	F(1,439) = 143.67, p < .001 *
	Diagnosis BDII	n/a	n/a	−0.74 (0.57)	F(1,438) = 1.65, p = .200
	Lithium	n/a	n/a	0.80 (0.38)	F(1,437) = 4.50, p = .034 *
	Antipsychotic	n/a	n/a	−0.90 (0.38)	F(1,438) = 5.61, p = .018 *
	Anticonvulsant	n/a	n/a	−0.79 (0.40)	F(1,438) = 3.84, p = .051
	Antidepressant	n/a	n/a	0.20 (0.37)	F(1,436) = 0.28, p = .597
	Age of onset	n/a	n/a	0.03 (0.02)	F(1,437) = 2.29, p = .131
	Psychosis	n/a	n/a	0.44 (0.46)	F(1,439) = 0.92, p = .338
Cortical surface area	Group (BD)	−0.03 (0.22)	F(1,2431) = 0.02, p = .904	n/a	n/a
	BMI	−0.10 (0.18)	F(1,2424) = 0.35, p = .554	0.41 (0.38)	F(1,439) = 1.20, p = .274
	Sex (F)	−5.77 (0.18)	F(1,2421) = 980.07, p = .000*	−6.40 (0.44)	F(1,436) = 210.13, p < .001*
	Age	−1.53 (0.13)	F(1,2430) = 139.82, p = .000*	−2.49 (0.36)	F(1,440) = 47.03, p < .001*
	Diagnosis BDII	n/a	n/a	−0.81 (0.74)	F(1,441) = 1.19, p = .276
	Lithium	n/a	n/a	1.55 (0.48)	F(1,440) = 10.23, p = .001*
	Antipsychotic	n/a	n/a	−0.36 (0.49)	F(1,441) = 0.54, p = .463
	Anticonvulsant	n/a	n/a	0.56 (0.52)	F(1,441) = 1.16, p = .282
	Antidepressant	n/a	n/a	0.33 (0.48)	F(1,437) = 0.47, p = .495
	Age of onset	n/a	n/a	0.03 (0.03)	F(1,441) = 1.01, p = .315
	Psychosis	n/a	n/a	0.34 (0.59)	F(1,440) = 0.33, p = .565

• Note: Asterisks indicate significant associations (*p < .05).

For surface area, only sex, age, and Li treatment were associated with the first component scores for surface area, see Table 2. Females, relative to males, and older participants showed significantly smaller surface area, while those prescribed lithium showed larger surface area.

4 DISCUSSION

In this study, the first of 68 total PCs accounted for 42.7% of variance in cortical thickness. The first PC, which indexed a greater cortical thickness across all 68 cortical regions, was negatively associated with BD, BMI, antipsychotic medications, and age and positively associated with Li treatment. These associations between the first PC and cortical thickness closely mirrored the associations found when applying clustering to the same sample, where the cluster with lower cortical thickness was also associated with diagnosis of BD, higher BMI, and older age, and the cluster with higher cortical thickness was associated with Li treatment (McWhinney, Abé, et al., 2022). Only PCA, not clustering detected links with antipsychotic medications. Also, when directly compared, PCA outperformed clustering as predictors of clinical and demographic variables. When we applied the PCA to the previously unseen replication sample collected from additional ENIGMA-BD working group sites, on which the model was not trained, we found the same patterns of associations with diagnosis and age, even though the sample was almost 90% smaller. The same pattern of brain changes detected in BD was also associated with the diagnosis of schizophrenia. Similar to previous large-scale studies, surface area was not associated with diagnosis of BD or BMI (Hibar et al., 2018; McWhinney, Abé, et al., 2022; McWhinney, Brosch, et al., 2022; McWhinney et al., 2023). The different system-level correlates of CT and SA in our and previous studies further support the practice of keeping these measures separate.

We directly compared PCA and clustering in the same dataset. Both methods detected similar associations with system-level variables, including diagnosis, BMI, age, and Li exposure. However, PCA outperformed clustering in terms of model fit and sensitivity. We suspect there are systematic reasons for this. Even if there are no clearly defined discontinuities/clusters in the data, clustering would segment continuous distribution of findings into several parts. Similar categorization of a continuous range of values necessarily results in a loss of statistical power, as very alike individuals are considered distinct when on opposing sides of a clustering threshold. That is, clustering does not encode the distance between individuals in the multidimensional space where those assigned to the same cluster still differ from one another. Similarly, as there are no strict boundaries, individuals assigned to different clusters may be very similar to one another. In contrast, PCA does not need arbitrary criteria to delineate patterns and find orthogonal effects in the dataset. Within the same component, we were able to maintain the strength of the associations and distance between individuals. For both reasons, PCA should systematically outperform clustering when there are no clearly defined groups of individuals and indeed that was the case in our study.

While our study replicates previous findings of negative associations between cortical thickness and diagnosis of BD in mass univariate analyses (Hibar et al., 2018; McWhinney, Abé, et al., 2021; McWhinney, Brosch, et al., 2022; van Erp et al., 2018), the effect size for association between first PC (d = 0.33) and diagnosis of BD was stronger than associations between individual ROIs and the diagnosis of BD in previous ENIGMA studies (d = 0.015–0.29) (Hibar et al., 2018) and instead of running one model per region, we captured covariance across all regions by a single number. Even in mass univariate analyses associations between clinical variables and brain structure are evident across many regions and are not isolated to a single ROI (Hibar et al., 2018; McWhinney, Abé, et al., 2021; McWhinney, Brosch, et al., 2022; van Erp et al., 2018). The lesion model, where changes in a single region are necessary and sufficient to cause an illness, clearly does not apply to a complex disorder such as BD (Hibar et al., 2018; Reddan et al., 2017). Consequently, looking at distributed effects across groups of regions should be more informative than looking at individual regions. All in all, it is encouraging that our findings converge with these theoretical expectations and suggest that multivariate analyses are better suited to studying complex neuropsychiatric disorders than mass univariate ones.

While the second component for cortical thickness explained approximately 12% of variance, it was associated predominantly with research site; with the outlying site removed, the second component explained only 5% of the variance, see supplement. It is interesting that the first and second components, which are necessarily orthogonal to one another, were each predominantly associated with clinical/anthropometric variables, or research site, respectively. These differences need to be replicated in other studies, but they may represent a more generalizable pattern. After all, correlates of demographic and clinical variables are consistent in the same direction (i.e. negatively associated with cortical thickness). In contrast, the variations related to research sites may be less consistent and less predictable. Consequently, they may fall into separate components. Interestingly when we applied a different method of removing the site effects, i.e. ComBat, the associations between the first PC and clinical/demographic variables remained identical to the results obtained without ComBat. Even without removal of site effects from the raw data, by identifying orthogonal components, PCA may implicitly separate the site effects from the more predictable/systematic biological effects. If confirmed in future studies, this would be a major advantage of PCA.

The same distributed pattern of brain regions was associated with each of the system-level variables. When we applied the PCA projections to the replication sample, we found associations with similar system-level variables, consistent with observations by others (Cao et al., 2023). In fact, the same patterns were also associated with the diagnosis of schizophrenia. This is in keeping with other large-scale studies or meta-analyses, which have also demonstrated that there is a common, non-specific pattern of case-control differences across major psychiatric disorders (Goodkind et al., 2015; Hettwer et al., 2022; Matsumoto et al., 2023), with highly correlated neurostructural abnormalities between BD and schizophrenia (Opel et al., 2020) and with PCA detecting a profile of shared cortical thickness differences across 6 major psychiatric disorders, which explained 48% of variance (Writing Committee for the Attention-Deficit/Hyperactivity Disorder et al., 2021). Some have argued that there may be more cause-specific alterations after removing this non-specific pattern of the first component (Cao et al., 2023), though with exception of the second component for cortical thickness, subsequent components explained only a small fraction of variance (i.e. typically less than 5% each). All in all, this study contributed to the growing body of evidence for lack of specificity of associations between some key clinical and demographic factors and patterns of brain alterations, so-called neural P factor (Sprooten et al., 2022). Many different variables, including age, obesity, psychiatric diagnoses may be negatively associated with cortical thickness across a wide range of cortical regions.

When testing associations with BMI and group, the first PC explained 28.6% more variance than average cortical thickness. Principal component analysis, which accounts for the regional distribution of effects, was better than simple average cortical thickness, which collapses information across all regions. At the same time, the PC1 was highly correlated with average cortical thickness. While there are regional effects and accounting for these improves the fit of the model, the system-level variables are associated with some level of cortical thinning across most regions. This may further contribute to the lack of specificity of the system level to brain associations and is consistent with findings from another study indicating that accuracy of ML classifier comparing controls versus BD versus schizophrenia was strongly dependent on global grey matter measures (Schwarz et al., 2019).

Our findings confirmed that BD is characterized by diffuse regional structural brain alterations, specifically lower cortical thickness. The fact that these alterations are so diffuse as to resemble global atrophy is interesting. We can only speculate about why this would occur. First of all, most pathologies will result in atrophy, i.e. thinning of the cortex. Second of all, changes in one region are likely to propagate through the network, thus eventually involving more related regions. Thirdly, perturbations within one network are going to propagate to other networks, thus involving even more regions eventually to the point of resembling a global atrophy. These mechanisms could explain why so many regions are correlated across individuals, why the association with the first component is uniformly in one direction, and even why different predictors (age, BMI, BD, schizophrenia, medications) are all associated with the same global pattern.

The advantages of this study include the large sample size, the validation of the most prominent findings in a replication sample, and the multivariate approach which improved statistical sensitivity. Importantly, we were able to directly compare results between two unique multivariate analyses in the same sample: clustering and PCA.

The multi-site nature of the study is a limitation that complicates the data analyses and interpretation of the results. At the same time, our findings suggest that PCA may code site in separate PCs from the effects of clinical/anthropometric variables. Along similar lines, the sample contained a broad representation of BD, including individuals with BDII and we also included a sample of people with schizophrenia. In order to test how well a method captures relevant clinical links, we need a representative/generalizable snapshot of the disorder, which is not restricted to one subtype and ideally also includes representation of other diagnoses. More detailed clinical or biological markers beyond those analyzed were not broadly available throughout the ENIGMA-BD working group. While it is possible that associations with other clinical variables would show different patterns, the consistent and replicated nature of the direction of associations and spatial distribution of networks suggest this is unlikely. As these were independently collected datasets, not a centralized single study, we did not have access to raw, whole-brain data.

We performed these analyses on derived estimates of cortical thickness and surface area, but we cannot generalize our findings to other measures, which may show different patterns of associations. There are other methods within this broad category, which may be of specific use for example when attempting a multimodal data fusion, i.e. FLICA. However, comparing different methods within the broad group of linear representation of data was not our goal and would likely only lead to incremental gains, if any. We did not test other multivariate techniques including ML methods, which are difficult to interpret. Our interest was in applying traditional methods which may also be applied in smaller samples and in a wide range of situations, including clinical settings. Due to simplicity and straightforward application of the linear methods, they may be a method of choice, especially where the size/structure of the dataset would not allow for proper ML.

5 CONCLUSION

In this study, we confirmed that cortical thickness in widespread regional networks as determined by PCA is negatively associated with relevant clinical and demographic variables, including diagnosis, age, BMI, and treatment with antipsychotic medications or lithium. The action of these factors on a widespread network suggests that conceptualizing or studying their effects on individual regions may be misleading. In addition, significant associations of many different system-level variables with the same network suggest a lack of specificity to individual clinical and demographic factors. While there may be general agreement among multivariate techniques on these associations, PCA outperformed clustering and may better fit the nature of brain imaging data in SMI. More broadly, the results have demonstrated that representing data as a linear combination of features is superior to clustering when investigating links between brain and system measures in SMI. This work could help researchers make informed decisions about which methods to use and could save them from applying an ill-fitting method, when a simpler, more reproducible one is available.

FUNDING INFORMATION

This work is also part of the German multicenter consortium “Neurobiology of Affective Disorders. A translational perspective on brain structure and function,” funded by the German Research Foundation (Deutsche Forschungsgemeinschaft DFG; Forschungsgruppe/Research Unit FOR2107). Principal investigators (PIs) with respective areas of responsibility in the FOR2107 consortium are as follows: Work Package WP1, FOR2107cohort and brain imaging: TK (speaker FOR2107; DFG grant numbers KI588/14-1, KI588/14-2, KI588/20-1, KI588/22-1), UD (co-speaker FOR2107; DA 1151/5-1, DA 1151/5-2), AK (KR 3822/5-1, KR 3822/7-2), IN (NE 2254/1-1, NE 2254/2-1, NE 2254/3-1, and NE 2254/4-1), BS (STR 1146/18-1), CK (KO 4291/3-1). Further support from the German sites was provided by MNC and FOR2107-Muenster: This work was funded by the German Research Foundation (SFB-TRR58, Project C09 to UD) and the Interdisciplinary Center for Clinical Research (IZKF) of the medical faculty of Münster (grant Dan3/012/17 to UD and grant SEED11/18 to NO); FOR2107-Muenster: This work was supported by grants from the Interdisciplinary Center for Clinical Research (IZKF) of the medical faculty of Münster (grant MzH 3/020/20 to TH) and the German Research Foundation (DFG grants HA7070/2-2, HA7070/3, HA7070/4 to TH, CRC1457/A7 to RN). The Medellin studies (GIPSI) were supported by the PRISMA UNION TEMPORAL (UNIVERSIDAD DE ANTIOQUIA/HOSPITAL SAN VICENTE FUNDACIÓN), Colciencias-INVITACIÓN 990 de 3 de agosto de 2017, Codigo 99059634. The San Raffaele site was supported by the Italian Ministry of Health RF-2018-12367789 project. The University of Galway research was supported by the Irish Research Council (IRC) Postgraduate Scholarship, Ireland awarded to LN and to GM, and by the Health Research Board (HRA-POR-324) awarded to DMC and (HRA_POR/2011/100) awarded to CMcD. We thank the participants and the support of the Welcome-Trust HRB Clinical Research Facility and the Centre for Advanced Medical Imaging, St. James Hospital, Dublin, Ireland. JS and RTK received support from the William K. Warren Foundation National Institute of Mental Health (R21MH113871); JS also acknowledges the National Institute of General Medical Sciences (P20GM121312). MB is supported by a NHMRC Senior Principal Research Fellowship and Leadership 3 Investigator grant (1156072 and 2017131). JH and EB were supported by the Czech Health Research Council Project No. NU21-08-00432 and by ERDF-Project Brain dynamics, No. CZ.02.01.01/00/22_008/0004643. IHG received support from the National Institute of Mental Health (R37MH101495). This study was also funded by EU-FP7-HEALTH-222963 “MOODIN-FLAME” and EU-FP7-PEOPLE-286334 “PSYCHAID.” The Barcelona group would like to thank CIBERSAM (EPC) and the Instituto de Salud Carlos III (PI18/00877 and PI19/00394) for their support. This work was supported by the Singapore Bioimaging Consortium (RP C009/2006) research grant awarded to K.S. The CIAM group (FMH—PI) was supported by the University Research Committee, University of Cape Town, and South African funding bodies National Research Foundation and Medical Research Council; DJS from CIAM was supported by the SAMRC. FS from CIAM was supported by a Brain-Behavior Unit Postdoctoral Research Fellowship. The Sydney studies were supported by the Australian National Health and Medical Research Council (NHMRC) Program Grant 1037196, Investigator Grants 1176716 (PRS) and 1177991 (PBM), Project Grants 1063960 and 1066177, the Lansdowne Foundation, Good Talk and Keith Pettigrew Family; as well as the Janette Mary O'Neil Research Fellowship to JMF. The study was also supported by NIMH grant number: R01 MH090553(to RAO). Funding for the Oslo-Malt cohort was provided by the South Eastern Norway Regional Health Authority (2015-078), the Ebbe Frøland Foundation, and a research grant from Mrs. Throne-Holst. EV acknowledges the support of the Spanish Ministry of Science and Innovation (PI15/00283, PI18/00805) integrated into the Plan Nacional de I + D + I and co-financed by the ISCIII-Subdirección General de Evaluación and the Fondo Europeo de Desarrollo Regional (FEDER); the Instituto de Salud Carlos III; the CIBER of Mental Health (CIBERSAM); the Secretaria d'Universitats i Recerca del Departament d'Economia i Coneixement (2017 SGR 1365), the CERCA Programme, and the Departament de Salut de la Generalitat de Catalunya for the PERIS grant SLT006/17/00357. Lastly, this study was supported by the Canadian Institutes of Health Research (103703, 106469, and 142255, 180449, 186254), Nova Scotia Health Research Foundation, Dalhousie Clinical Research Scholarship to TH, Brain & Behavior Research Foundation (formerly NARSAD); 2007 Young Investigator and 2015 Independent Investigator Awards to TH. The Minnesota sites were supported by the National Institutes of Health (U01MH108150 to SS; K01MH093621 to SU), the Center for Magnetic Resonance Research (P41 EB027061; 1S10OD017974), a Merit Review Award (#I01CX000227 to SS) from the United States (U.S.) Department of Veterans Affairs Clinical Science Research and Development Service, and Minneapolis VA Health Care System resources to SU and SS. The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of the Department of Veterans Affairs or the United States government. LTE was supported by NIH MH083968 and Desert-Pacific Mental Illness Research Education and Clinical Center.

EV thanks the support by CIBER-Consorcio Centro de Investigación Biomédica en Red-(CB07/09/0004), Instituto de Salud Carlos III, Spanish Ministry of Science and Innovation and grants PI18/00805 and PI21/00787, integrated into the Plan Nacional de I + D + I and co-financed by the ISCIII-Subdirección General de Evaluación and the Fondo Europeo de Desarrollo Regional (FEDER); the Instituto de Salud Carlos III; the Secretaria d'Universitats i Recerca del Departament d'Economia i Coneixement (2021 SGR 01358), the CERCA Programme, and the Departament de Salut de la Generalitat de Catalunya for the PERIS grant SLT006/17/00357. Thanks also for the support of the European Union Horizon 2020 research and innovation program (EU.3.1.1. Understanding health, wellbeing, and disease: Grant No 754907 and EU.3.1.3. Treating and managing disease: Grant No 945151).

JR thanks the support by the Spanish Ministry of Science and Innovation. Instituto de Salud Carlos III (PI19/00394 and PI22/00261), integrated into the Plan Nacional de I + D + I and co-financed by ERDF Funds from the European Commission (“A Way of Making Europe”), CIBERSAM, and the Secretaria d'Universitats i Recerca, Departament d'Economia i Coneixement and Departament de Salut (2021 SGR 01128).

CONFLICT OF INTEREST STATEMENT

PMT & CRKC received a grant from Biogen, Inc., for research unrelated to this manuscript. DJS has received research grants and/or consultancy honoraria from Lundbeck and Sun. LNY has received speaking/consulting fees and/or research grants from Abbvie, Alkermes, Allergan, AstraZeneca, CANMAT, CIHR, Dainippon Sumitomo Pharma, Janssen, Lundbeck, Otsuka, Sunovion, and Teva. TE received speaker's honoraria from Lundbeck and Janssen Cilag and is a consultant to Sumitomo Pharma America. Thanks also for the support of the European Union Horizon 2020 research and innovation program (EU.3.1.1. Understanding health, wellbeing, and disease: Grant No 754907 and EU.3.1.3. Treating and managing disease: Grant No 945151). EV has received grants and served as consultant, advisor, or CME speaker for the following entities (unrelated to the present work): AB-Biotics, Abbott, Allergan, Angelini, Dainippon Sumitomo Pharma, Ferrer, Gedeon Richter, Janssen, Lundbeck, Otsuka, Sage, Sanofi-Aventis, and Takeda. PMT and CRKC have received partial research support from Biogen, Inc. (Boston, USA) for work unrelated to the topic of this manuscript. EV has received grants and served as consultant, advisor, or CME speaker for the following entities: AB-Biotics, AbbVie, Adamed, Angelini, Biogen, Biohaven, Boehringer-Ingelheim, Celon Pharma, Compass, Dainippon Sumitomo Pharma, Ethypharm, Ferrer, Gedeon Richter, GH Research, Glaxo-Smith Kline, HMNC, Idorsia, Johnson & Johnson, Lundbeck, Medincell, Merck, Novartis, Orion Corporation, Organon, Otsuka, Roche, Rovi, Sage, Sanofi-Aventis, Sunovion, Takeda, and Viatris, outside the submitted work. Yatham reports grants from Abbvie and Dainippon Sumitomo, and served as an advisor or consultant or speaker to JAMA Pharma, Intracellular Therapies, Merck, Allergan, GSK, Gedeon Richter, Sanofi, Sunovion, and Alkermes, outside the submitted work.