|Home | About | Journals | Submit | Contact Us | Français|
The complex molecular etiology of psychosis in schizophrenia (SZ) and psychotic bipolar disorder (PBP) is not well defined, presumably due to their multifactorial genetic architecture. Neurobiological correlates of psychosis can be identified through genetic associations of intermediate phenotypes such as event-related potential (ERP) from auditory paired stimulus processing (APSP). Various ERP components of APSP are heritable and aberrant in SZ, PBP and their relatives, but their multivariate genetic factors are less explored.
We investigated the multivariate polygenic association of ERP from 64-sensor auditory paired stimulus data in 149 SZ, 209 PBP probands, and 99 healthy individuals from the multisite Bipolar-Schizophrenia Network on Intermediate Phenotypes study. Multivariate association of 64-channel APSP waveforms with a subset of 16 999 single nucleotide polymorphisms (SNPs) (reduced from 1 million SNP array) was examined using parallel independent component analysis (Para-ICA). Biological pathways associated with the genes were assessed using enrichment-based analysis tools.
Para-ICA identified 2 ERP components, of which one was significantly correlated with a genetic network comprising multiple linearly coupled gene variants that explained ~4% of the ERP phenotype variance. Enrichment analysis revealed epidermal growth factor, endocannabinoid signaling, glutamatergic synapse and maltohexaose transport associated with P2 component of the N1-P2 ERP waveform. This ERP component also showed deficits in SZ and PBP.
Aberrant P2 component in psychosis was associated with gene networks regulating several fundamental biologic functions, either general or specific to nervous system development. The pathways and processes underlying the gene clusters play a crucial role in brain function, plausibly implicated in psychosis.
Growing evidence supports common genetic mechanisms underlying the psychosis dimension in schizophrenia (SZ), schizoaffective (SZA), and psychotic bipolar disorder (PBP).1–3 Numerous genetic markers have been implicated in the etiology of these disorders,4 yet none of them prove to have a high penetrance,5 consistent with a polygenic common disease common variant (CDCV) model comprising multiple interacting genes contributing to a cumulative effect, distinct from the weak individual gene effect in disease risk or association.6 Further, the underlying pathways and mechanisms of the genes mediating the neurophysiological abnormalities and clinical symptoms are not well ascertained,7,8 and their polygenic effects are relatively unstudied. An effective strategy for gene and disease target discovery underlying complex neuropsychiatric diseases is to employ endophenotypes that are heritable state-independent biological measures associated with illness due to shared genetic influences. The concept of endophenotype in psychiatric genetics has elicited increased attention in gene characterization due to the presumed tractable genetic framework and close proximity to the level of gene action,9 as opposed to the complex clinical disease.
One such endophenotype is the event-related potential (ERP) measured via noninvasive paired click auditory sensory processing (auditory paired stimulus processing [APSP]) paradigm. Auditory sensory gating is assessed with the paired stimulus paradigm, involving presentation of 2 (S1, S2) closely separated identical sounds (500ms inter-stimulus interval), with long intervals (6–10 s) between stimulus pairs.10,11 The earliest component of the mid-latency ERPs is the P50, a positive wave peaking at ~50ms (between 35 and 85ms) following presentation of each stimulus.12 Aberrant P50 gating is a key biomarker marker of SZ and PBP13–15 and also abnormal in unaffected relatives.16 Informative measurements in paired-stimuli paradigms are not limited to P50 peaks. ERPs derived from APSP paradigms also comprise complex cascade of mid/late latency including N1, the negative voltage ERP peaking approximately 100ms and P2, the positive ERP peaking approximately 200ms after S1 onset and sustained deflections that carry unique, substantial, and complementary elements depicting SZ and PBP pathophysiology.10,17,18 The N1 component is also compromised in unaffected SZ relatives.19 P50, N1, and P2 components are associated with various attentive stages of information processing.20,21 Less traditional ERP measurements from APSP, such as a pre-S2 negativity, have been shown to capture heritable disease heterogeneity in a large sample of SZ and PBP proband and first-degree relatives.22 That study did not include SZA probands, thus distinctions between SZ and PBP from SZA probands were not investigated. Despite the lack of clear pathophysiological boundary between SZ and SZA, and the clinical heterogeneity manifested in SZA disorder (includes depressive and manic subtypes), an earlier study examined auditory P300 ERP in both these probands and reported that SZA did not exhibit P300 amplitude deficits observed in SZ.23 A recent report24 showed that P300 feature of response inhibition did not differentiate SZA from SZ and controls. Additionally, a previous study employing electrophysiological indices demonstrated lack of diagnostic specificity in comparing SZA to SZ and PBP but revealed that such measures strongly differentiated SZ subtypes.25 A detailed review of electrophysiological studies comparing SZA to other controls and probands can be found in ref.26. The inclusion of SZA probands in either the SZ or PBP group has been debated due to clinical heterogeneity and overlapping symptoms.26 The current study analyzed the entire ERP waveform derived from APSP including all 64 sensors across the entire scalp. Such a strategy offers a comprehensive representation of the biomarkers compared to a single biomarker based genetic studies. In this study, we empirically assessed time-voltage ERP from subset of probands in conjunction with genetic data, different from prior study22 examining both time-voltage and time-frequency activity across probands and relatives. Although significant familiality (h 2 0.2–0.5) was noted in that study no unaffected relatives’ effects were present.
Although few reports indicate P50 impairments to be associated with dysfunctions in various neurotransmitter systems,27–29 the neurobiological correlates of other APSP components are known and not well characterized. As an endophenotype, heritability of APSP components has been substantiated by several twin studies.19,30–32 Genetic linkage studies have proposed various risk loci and candidate genes for P50 abnormalities in SZ and PBP including the 15q13-14 (the site of alpha-7 nicotinic acetylcholine receptor gene CHRNA7),33,34 22q11-q12,35 18q21,36 COMT,37 and CACNA1C.38 However, these findings have not been consistently replicated.39 One prior study examined genetic associations of N1 and P2 ERP components but yielded no significant linkage.40 To our knowledge, no prior studies have extensively evaluated the genetic correlates of the N1 and P2 components in psychosis.
Genome-wide association (GWA) of single nucleotide polymorphisms (SNPs) or pedigree-based linkage analysis is the primary tool for identifying complex traits in most genetic studies. The univariate strategy of GWA examines a SNP with single phenotype of interest. Thus, failing to account for the multifactorial model,41 where linear and epistatic interaction between multiple SNPs confers risk for complex psychiatric diseases.42 For instance, because most of the common SNP variants usually confer small increments in risk of the disease, they often do not reach genome-wide significance due to stringent multiple comparison correction for numerous SNP variants in the GWA model. Thus, large study samples are required to achieve genome wide significance. To mitigate these problems and improve the predictive power of complex phenotypes in moderate sized samples, novel statistically efficient data-driven multivariate association approaches based on parallel independent component analysis (Para-ICA)43,44 are employed to model the presumed polygenic architecture of complex psychiatric illnesses.45 Such methods have been successfully used to identify gene correlates implicated in Alzheimer’s disease46 and psychoses.47–49 In multivariate model, the aggregate constituent from linearly coupled gene variants is linked (linear additive model) to complex phenotypes including multiple biological measures (as opposed to single feature in GWA). Under the multifactorial assumption, several gene variants mediate psychotic disorders in concordance with the notion that multiple complex neurobiological mechanisms underlie the disease etiology.
The main goals in this study are to (1) identify prior psychosis candidate genes and novel genes associated with auditory processing abnormalities in psychosis. (2) Examine gene networks and associated molecular and biological attributes collectively influencing distinctive auditory processing features in SZ and PBP. In line with prior evidence, we hypothesized that ERP abnormalities in APSP would be influenced by genes involved in axon guidance and synaptic transmission known to be associated with psychosis.
The genotype–phenotype multivariate association was evaluated using multimodal data fusion (integration of data from multiple modalities to leverage the joint or combined information for better characterization of brain activity) of SNP variants and ERP data from APSP pooled across SZ, PBP, and healthy comparison (HC) individuals from the Bipolar-Schizophrenia Network on Intermediate Phenotypes (BSNIP) study.
We examined 457 subjects including 99 HC, 149 SZ, and 209 PBP probands. Study participants in this sample were selected such that each had both genotype data (N = 620 subjects inclusive of SZ, SZA and PBP probands and HC only) and had participated in APSP paradigm (N 1600 subjects comprising all probands, relatives and HC). A subset of these APSP ERP data (N 1100) were previously published elsewhere.22 Thus the sample comprised an intersection of both subsets (N = 620 and 1600 subjects), which were part of the overall 2441-person cohort from the BSNIP study (refer to ref.50 for complete details including participant recruitment and exclusion criteria). The subjects were aged 15–65 years and recruited (see supplementary material) from the 5 participating BSNIP sites. Probands were on stable doses of psychotropic medication (supplementary table ST1) for ≥4 weeks. The study was explained to all subjects and institutional review board approved written informed consent was obtained at the respective participating sites. Detailed demographic characteristics and medication information for the study sample are listed in table 1 and supplementary table ST1, respectively.
Electroencephalogram (EEG) data were collected at each site from identical Neuroscan equipment equipped with 64 Ag/AgCl electrodes (Quik-Cap, Compumedrics), while subjects listened to 150 binaural broadband auditory stimuli pairs (4ms duration at 75 dB), separated by an average of 9.5 seconds, with 500ms between stimuli in a pair (see supplementary material). Electrode positions were defined according to the International 10-10 EEG system, with the inclusion of mastoids and CP1/2 locations to provide greater signal sampling below the cantho-meatal line. Nose served as a reference with a forehead ground. Electrode impedances were maintained below 5 KΩ throughout the procedure. EEG recordings were amplified with a gain of 12 500 and digitized at a sampling rate of 1000 Hz, using Neuroscan Acquire and SynAmps2 recording systems (Compumedics Neuroscan).
DNA was extracted using blood samples collected from participants. Approximately 1 million (1 140 419) target SNP variants across the whole genome were genotyped using the Illumina Human Omni1-Quad chip and BeadArray platform at Genomas Inc, Hartford Hospital.
Raw EEG data were inspected and adjusted (see supplementary figure SF1) for bad electrodes and artifacts (<5% for any subject), using spherical spline interpolation (BESA 5.3; MEGIS Software). Data were converted to an average reference and digitally band-pass filtered from 0.5–55 Hz (zero phase filter; rolloff: 6 and 48 dB/octave, respectively). The current filter settings are suitable for deriving the N1 and P2 ERP components, while they were suboptimal to extract the P50 component.51 Refer to supplementary material for additional details on ERP data processing.
Genotyped SNP data typically represented as homozygous (AA, BB) and heterozygous (AB or BA) were converted to numerical format by additive coding for the number of minor alleles (AA = 0, AB = 1, and BB = 2, assuming B is minor allele). SNP data were subjected to quality control process outlined in ref.52 using Plink (see supplementary figure SF1 and supplementary text for processing details) and stratification bias (see Q-Q plot in supplementary figure SF2) was corrected using principal component analysis (PCA)53 (using custom Matlab scripts). Data were reduced from a ~106 to 16 999 SNPs by selecting those significant at P < .05 in either of the proband (SZ, PBP) vs control logistic regression.
Data were constructed as a matrix of (1) subjects by SNPs (457×16 999), (2) subjects by ERP waveforms (457×80 000 [64 channels × 1250 time points of ERP data]) and were reduced to 12 and 2 hidden components or factors, respectively (see supplementary text). ERP and SNP data were simultaneously processed by Para-ICA43,44 (see supplementary figure SF3). The original ERP data,22 were processed using principal component analysis (see supplementary text for PCA vs ICA differences). A leave-one-out cross validation was conducted using the same parameters used in the original run to assess the reliability of SNP and ERP components.
The strength of association was evaluated using correlations between SNP and ERP loading coefficients (LC). The effects of confounding demographic factors including age, sex and site on LCs were accounted using partial-correlation. Correlations for all component pair combinations (2×12 = 24) were evaluated and significance levels adjusted using Bonferroni multiple comparison correction for number of combinations (P < .05/24). Supplementary analyses including partial-correlation accounting for group variations, multiple regression analysis and correlation with symptoms are presented in the supplementary material. ERP and SNP LCs of significant component pairs were examined between groups using pairwise t tests. The SNPs and ERP features in each component contributing to the association were identified using a threshold of |Z| = 2.5.
Genes representing SNPs selected from the genetic network were entered into the GeneGo MetaCore software, version 6.17 (https://portal.genego.com) and the ConsensusPathDB database (CPDB), release 28 (http://consensuspathdb.org)54 (includes KEGG and Reactome resources from CPDB) to examine the biological pathways/mechanisms underlying sensory gating abnormalities (see supplementary material).
We identified 1 significant gene-ERP linkage. Of the 2 ERP components (E1-E2) derived from Para-ICA (shown in supplementary figure SF4), E1 was significantly correlated with G2 (r = .19, P = .00004, N = 457) (after multiple comparison correction). The linkage between ERP and SNP LC was assessed using partial-correlation after adjusting for the effects of demographic variables (see supplementary table ST2 for their effects on LC). The association between E1-G2 was significant when both proband groups and HC were combined as a single group (figure 1); however no significant association was observed within each group (SZ, PBP, and HC) separately (see supplementary figure SF5), likely due to constrained LC variability (LC range) within each diagnostic group. The component pair E2-G3 did not yield a significant linkage after Bonferroni adjustment (r = −.09, P = .07, N = 457). Refer to supplementary material for results from the partial-correlation after adjusting for group variations (supplementary figure SF6) and multiple regression analysis.
E1 and E2 comprised N1-P2 wave complex with a dominant P2 and secondary N1 feature, respectively. The ERP components are plotted at peak location Fcz and the topography for E1 and E2 was mapped using the peak amplitude in the P2 (150–250ms) and N1 time range (75–130ms) in response to stimulus S1, respectively. The topographies for both components revealed a fronto-central distribution. E1 was positively correlated with G2 (see figure 1 for scatter plot), reflecting increased linear combination of minor allelic variation related to increased E1, respectively. The genetic network G2 comprised 281 SNPs/170 unique genes. The top 20 genes from the genetic network G2, their functional properties and their known associations with neuropsychiatric diseases are summarized in table 2 and supplementary table ST3, respectively. No significant association was found between ERP LC and symptom measures including Positive and Negative Syndrome Scale (PANSS) positive (r = −.02, P = .6), PANSS negative (r = −.027, P = .6), PANSS general (r = .037, P = .48), MADRS (r = −.015, P = .77), Young mania (r = −.07, P = .18), and SBS ratings (r = .035, P = .51). Group comparisons on ERP and SNP LC revealed significant differences between HC and both SZ and PBP proband groups for E1-G2 pair (see figure 1 and supplementary figure SF7). A P2 (response to S1) deficit was observed in both proband groups (HC > SZ > PBP) compared to HC. Further, pairwise comparison revealed a trend toward greater P2 deficit in PBP compared to SZ probands [t(1.69), P < .08; see figure 1]. Reliability evaluated using leave-one-out cross-validation revealed an average within modality correlation > 0.9, indicating stable ERP and SNP components.
The significant pathways found in GeneGo and consensus pathway data base (CPDB) databases are listed in supplementary tables ST4 and ST5, respectively. G2 network (related to E1 network) was significantly enriched with genes belonging to retrograde endocannabinoid signaling and glutamatergic synapse (see figure 2) in CPDB. Significant pathway maps in GeneGo included epidermal growth factor receptor (EGFR), immune response and development slit-robo signaling. Several process networks associated with G2 included GnRH, cholecystokinin, gonadotropin, WNT signaling and development neurogenesis (axon guidance and synaptogenesis). Prominent metabolic networks included maltohexaose pathway and transport.
The complex genetic architecture underlying psychosis in SZ and PBP is largely unknown. Etiology can be examined by probing intermediate phenotypes such as ERPs derived from APSP, a biological measure presumably closer to gene action than clinical phenomenology. Various features from APSP exhibit significant heritability in family-based studies30 and less densely related individuals,22 suggesting genetic control over ERP components. We hypothesized that multiple coordinated genes contribute to ERP abnormalities in APSP in SZ and PBP.
We observed deficits in E1 (dominant P2 response to S1) component in SZ and PBP, consistent with previous studies.17,22 P2 is related to filtering mechanisms associated with attention allocation.20 Although the neural basis of P2 is not well characterized, P2 deficits in patients may index disruptions of sensory registration of information allocated by stimuli and problems with auditory stimulus identification and classification.55
Para-ICA identified networks of synchronized genes, with each gene associated with a weight reflecting the individual contribution to the ERP-genetic linkage. It is important to note that both the P2 ERP and the related gene network showed pairwise group-differences in a similar ordinal fashion (HC > SZ > PBP), perhaps reflecting a common latent or hidden factor partly mediating the variability in both phenotype and genotype, further contributing to the diagnostic variations. This important finding revealed that a common source of influence may explain the existence of correlation between the ERP-gene networks that was influenced by diagnosis, plausibly providing insights towards search of trans-diagnostic psychosis dimensions. Variance in both ERP and genetic networks may thus influence the disease outcome. The genetic network identified in this study explained ~4% of the ERP phenotype variance, similar to other polygenic studies that have reported slightly higher explained variance than GWAS type studies1 (~2%). Thus, a majority of the phenotypic variability is still unresolved, likely due to the small contributions of each SNP toward the total variance. Additionally, environmental factors may influence the phenotype. We discuss the genetic basis of P2 abnormalities by examining the functionality of the top-ranked and repeatedly identified genes (multiple SNP occurrences), as well as the biological pathways enriched in the gene clusters.
The highest ranking gene in G2 was SCAI with 7 SNPs identified within the network. This gene encodes a protein regulating cell migration and function in the RhoA-Dia1 signal transduction pathway. The next ranking gene was PPP6C, encoding the catalytic subunit of protein phosphate regulating cell cycle progression. Although neither gene has no known specific central nervous system functions or neuropsychiatric disease association, cell cycle regulation plays a key postmitotic role in neuronal migration, axon pruning and synapse maturation.56 The CSMD1 gene (12 SNPs) encodes complement control-related protein suggested to inhibit the classical immune c3 activation,57 while also involved in neuronal wiring and altering synaptic strength.58 This gene plays an active role in immune system pathways and is associated with psychosis risk.59–61 ASIC2 encodes the degenerin/epithelial sodium channel associated with response to lithium treatment in PBP patients.62 SHANK2 encodes synaptic proteins that functions as molecular scaffolds in excitatory synapses implicated in autism risk.63 This gene is also involved in neuronal morphogenesis and regulates calcium signaling.64
Pathway enrichment analysis and gene network ontologies are ongoing fields of development.65 Although various gene annotation databases exist, these differ in underlying statistical approaches, thus presenting challenges in interpretation, and requiring refinement. However, such techniques are essential in identifying key biological substrates associated with key endophenotypes in complex psychiatric disorders. Numerous signaling pathways comprising endocannabinoid, EGF receptor, immune response (IL-15), tubby, developmental slit-robo influenced P2 in psychosis. Retrograde endocannabinoid signaling mediates synaptic plasticity in various brain regions.66 In the hippocampus, this pathway is triggered by NMDA receptor-mediated calcium entry into neurons.67 Endocannabinoid signaling regulates hippocampal GABA release68 and glutamatergic neurotransmission through D2 dopamine receptors.69,70 EGF receptor signaling has neurotrophic effects on developing dopaminergic neurons71 playing major role in neuronal proliferation and differentiation.72 One prior study reported abnormal EGF generation in chronic SZ tissues.73 EGFR (ErbB-1) is a member of the ErbB receptor tyrosine kinase proteins, and interacts with neuregulins, including NRG1, associated with SZ susceptibility.74
Additional signaling networks including gonadotropin-releasing hormone (GnRH), cholecystokinin and WNT were enriched in G2. GnRH is highly expressed in hippocampal75 and hypothalamic pyramidal neurons76 and contributes to pathology of SZ.77 Hypothalamic-pituitary-gonadal axis dysfunction is linked to affective arousal deficits in SZ.78 WNT signaling is associated with cortical development,79 hippocampal neurogenesis80 and synaptic plasticity.81 It is implicated in SZ through interaction with DISC182,83 and is found to be upregulated in bipolar disorder.84 The top metabolic networks including maltohexaose and maltopentaose, and 2 other carbohydrate metabolism networks related to glucose and sucrose were enriched in G2. Although the mechanism of metabolism in psychoses is unclear, prior evidence points to role of estrogen based glucose metabolism in SZ.85 Glucose metabolism interacts with dopamine signaling,86,87 a key network implicated in depression, SZ and PBP.
Validation of the current findings can be evaluated as percent overlap between genes\SNPs or pathways88 obtained from an independent sample, which was unavailable in this study. Several candidate genes (N = 13) and mechanisms reported in the current study have previously been implicated in psychosis. In particular, the large scale Psychiatric Genomics Consortium dataset identified immune and neuronal signaling pathways89 associated with SZ. Similarly, the Collaborative Study on Genetics (COGS) of SZ examined association of custom chip SNPs with various (N = 12) endophenotypes including P50 and identified axonal guidance and glutamate signaling90,91 pathways. Additional, pathway analysis studies revealed axon guidance,47,92 netrin signaling, neural cell adhesion process,93,94 immune response signaling,95 and glutamatergic synapse96 underlying psychosis. The individual genes identified using various BSNIP intermediate phenotypes may differ, as these measures assay differential neurophysiological aspects. Several neuronal processes (Venn diagram in figure 3) including neuron projection, differentiation, development, neurogenesis and axon guidance mediating various BSNIP EEG phenotypes comprising auditory oddball ERPs,49 resting EEG frequency activity97,98 and APSP ERPs in psychoses are comparable and similar, validating the current findings and our approach to combine functional and genetic data to unravel the complex etiology underlying these disorders.
Study limitations include: (1) Demographic variables including age and sex were unbalanced across data collection sites. (2) We grouped the SZA depressive- and manic-type probands into SZ and PBP, respectively. An ongoing debate exists on the inclusion of SZA in psychiatric nosology, perhaps, due to its clinical heterogeneity and lack of sufficient comparisons among SZA vs SZ and PBP. Several prior studies included SZA probands in either SZ or PBP categories, but there is no consensus on a best strategy; however a dimensional approach may be utilized in future studies to circumvent this problem. Thus, the current findings have to be interpreted with caution due to the boundaries we adopted in classifying the SZA probands. (3) Effects of sex and age were regressed through post hoc partial correlation between the loading coefficients of ERP and SNP modalities. (4) Another limitation was possible confounding effects of medications, which could not be controlled for, due to significant variation in the number of medications and duration of use (probands were using multiple medications making it difficult to dissect these into cells). (5) Relatives of probands were not included in the association analyses due to lack of genetic data. (6) Multivariate genetic analyses were confined to only subset of disease associated SNPs, excluding the possibility of identifying other key genetic networks. The current findings may be sensitive to the SNPs selected based on individual logistic regression analysis, comparing each proband group to HC. Future studies can investigate multivariate association of SNPs, selected from univariate analysis comparing all probands to HC, which would clarify the pathways common to psychosis across all probands. (7) Although population stratification effects were corrected using top 3 principal components (associated with race), sample admixture may have influenced the findings. Additionally, missing SNPs were imputed, which may introduce bias in association analysis. Hence we advise caution in interpreting these data. Future studies may employ sophisticated imputation techniques utilizing a reference haplotype to minimize imputation bias. (8) We were unable to verify the current findings due to the lack of replication sample. (9) Samples were too small to examine disorder specific effects.
Several studies have investigated GWA data for SZ and PBP, while some have conducted hypothesis-driven types of pathway analyses. In the current study, we were able to identify data-driven multivariate ERP-genetic associations by probing the auditory sensory system in psychotic probands via a noninvasive paired click paradigm. The genetic network identified in this study explained ~4% of the ERP phenotype variance. The ERP-genetic association is influenced by probands vs controls group difference; specifically PBP-HC group difference was highly prominent. The ERP-genetic association influenced by diagnosis may be of interest to clinical neuroscientists in the identification of trans-diagnostic dimensions of psychosis variance. The most significant pathways and processes that enriched the gene networks were related to several fundamental biologic functions, either general or specific to the nervous system, including neuronal differentiation, neurogenesis, glutamatergic, EGFR, endocannabinoid, and immune signaling. In summary, several genes and multiple biological pathways that have direct impact on brain function and development were found to be associated with P2 abnormalities in psychosis. The pathways identified in this study may be explored for therapeutic targets in future animal model studies. The present study did not examine the P50 ERP component; whose genetic basis will be explored in future studies. Future genetic studies can also examine epistasis among the identified genes.
Supplementary material is available at http://schizophreniabulletin.oxfordjournals.org.
National Institute of Mental Health (R01 MH077851 to G.D.P.; MH078113 to C.A.T.; MH077945 to M.S.K.; MH077862 to J.A.S.; MH077852 to Dr Gunavant Thaker). Additional support (2R44 MH075481 to G.R.; P20GM1034672 and R01EB006841 to V.D.C.).
We thank the study participants for their contributions.
J.A.S. has received support from Takeda, BMS, Lilly, Roche and Janssen. M.S.K. has received support from Sunovion. C.A.T. has received funding from Astellas, Eli Lilly, Intracellular Therapies, Lundback, and Pure Tech Ventures. Other authors declare no financial interest in relation to the work described in this manuscript other than the grant funding.