Multivariate genome-wide association study of sleep health demonstrates unity and diversity

Abstract

There has been a recent push to focus sleep research less on disordered sleep and more on the dimensional sleep health. Sleep health incorporates several dimensions of sleep: chronotype, efficiency, daytime alertness, duration, regularity, and satisfaction with sleep. A previous study demonstrated sleep health domains correlate only moderately with each other at the genomic level (|rGs| = 0.11–0.51) and show unique relationships with psychiatric domains (controlling for shared variances, duration, alertness, and non-insomnia independently related to a factor for internalizing psychopathology). Of the domains assessed, circadian preference was the least genetically correlated with all other facets of sleep health. This pattern is important because it suggests sleep health should be considered a multifaceted construct rather than a unitary construct. Prior genome-wide association studies (GWASs) have vastly increased our knowledge of the biological underpinnings of specific sleep traits but have only focused on univariate analyses. We present the first multivariate GWAS of sleep and circadian health (multivariate circadian preference, efficiency, and alertness factors, and three single-indicator factors of insomnia, duration, and regularity) using genomic structural equation modeling. We replicated loci found in prior sleep GWASs, but also discovered “novel” loci for each factor and found little evidence for genomic heterogeneity. While we saw overlapping genomic enrichment in subcortical brain regions and shared associations with external traits, much of the genetic architecture (loci, mapped genes, and enriched pathways) was diverse among sleep domains. These results confirm sleep health as a family of correlated but genetically distinct domains, which has important health implications.

Graphical Abstract

Graphical Abstract.

Statement of Significance.

We present the first multivariate genome-wide association studies of sleep and circadian health (multivariate circadian preference, efficiency, and alertness factors, and three single-indicator factors of insomnia, duration, and regularity) using genomic structural equation modeling. We found overlapping genomic enrichment in subcortical brain regions and shared associations with external traits, but that much of the genetic architecture (loci, mapped genes, and enriched pathways) was diverse among sleep domains.

Introduction

Good sleep is an integral part of daily functioning, physical, and mental health [1]. Indeed, there is a global impact associated with overall poor sleep health, as it relates to many negative health outcomes including diabetes, heart disease, stroke, and psychiatric disorders. Sleep health is a multi-dimensional construct that consists of several sleep domains, most often measures of quality, quantity, and timing of sleep [2]. Different domains of sleep and circadian health do not always correlate highly with each other, and relate differentially to many external outcomes such as mental health, chronic diseases, or general health [3–6]. The unique variance associated with each sleep health domain is key to understanding the etiology of sleep health and its nuanced relationships with other health problems. Here, we conduct several multivariate genome-wide association studies (GWASs) on sleep and circadian health and characterize the biological pathways and phenotypic correlates associated with six sleep domains.

Sleep health consists of chronotype, duration, daytime alertness, efficiency, satisfaction, and regularity of sleep. Studies of sleep health focus on these characteristics of healthy sleep rather than specific sleep disorders per se [2]. It is informative to study sleep from a positive health-based perspective because everyone can be measured on general, continuous sleep questions, whereas it can be challenging to obtain sufficient sample sizes of those who meet criteria for clinical sleep disorders [5]. Similarly, studying sleep on a continuum as a modifiable health behavior and identifying typical ranges and metrics of sleep can inform interventions or healthy benchmarks around the numerous health outcomes associated with poor sleep [7].

Studying sleep health from a genetic standpoint is equally important because many sleep problems have a substantial genetic component [8]. Several GWASs on sleep traits such as insomnia, chronotype, daytime sleepiness, sleep duration, and efficiency have found heritabilities ranging from 0.5% to 13.7%, and have discovered numerous genomic loci associated with these traits [9–13]. These GWASs have identified genes associated with individual sleep traits and genetic overlap with external traits such as psychopathology, cardiovascular diseases, and cognitive abilities. While these GWASs have been informative in understanding the genomic etiology of specific sleep traits, few have directly addressed the shared genomic liability across multiple sleep traits [14].

Transdiagnostic multivariate research has become increasingly popular as genomic studies have revealed widespread shared genetic liability among complex traits. The search for broader biological systems implicated across traits has the potential to be more informative than systems or genes pertaining to individual single traits [15]. Studying sleep health domains, as opposed to singular sleep disorders, can aid in informing treatment, interventions, and overall disease classification for a broader spectrum of sleep. Furthermore, the domains of sleep health are applicable to all sleep disorders (i.e. everyone with a sleep disorder can be measured on duration, regulation, or satisfaction of sleep) and do not occur in isolation [4]. Therefore, it can be a powerful approach to study multiple dimensions of sleep health in unison.

Previously, we used genomic structural equation modeling (Genomic SEM) [16] to test several structures and model the latent genetic structure of sleep health [6]. Using the covariance structure from 12 previously published GWAS summary statistics, we found the best-fitting model contained three multi-indicator sleep health factors: circadian preference (loadings for chronotype [10], most active 10 hours of the day [M10] [11], least active 5 hours of the day [L5] [11], and sleep midpoint [mid] [11]); Alertness (loadings for daytime sleepiness [13], napping [17], and diurnal inactivity [11]); Efficiency (loadings for sleep efficiency [11] and sleep episodes [11]); and three single-indicator factors: duration (loading for self-reported sleep duration [9]); non-insomnia (loading for reverse-coded insomnia [12]) and Regularity (loading for reverse-coded standard deviation of actigraphy-based sleep duration [11]).

Figure 1 depicts the sleep health model of this earlier study [6]. We found that these sleep and circadian health factors correlated weakly to moderately at the genetic level. Circadian preference was the least correlated with the other sleep health factors (rGs = |0.04–0.26|), while the other factors were more moderately correlated (rGs = |0.11–0.51|). Generally, these sleep health factors were negatively associated with psychopathology, and they predicted unique variance in latent internalizing, externalizing, psychosis thought disorders, and compulsive thought disorders factors. These patterns suggest that both the overlapping and unique genetic liability for sleep health domains are essential to furthering our knowledge of how poor sleep health manifests and understanding comorbidities.

Figure 1.

Panel A is the genomic structural equation model (SEM) of six correlated sleep health factors from Morrison et al. [6]. Factor correlations are shown in panel B. Boldface font for factor loadings indicates p < 0.05 [standard errors in brackets]. Boldface font for factor correlations indicates significance after false-discovery-rate correction.

The current study leverages both the previously published sleep trait GWASs and the multivariate structure identified by Morrison et al. [6] by performing a GWAS on the factors to provide insight into the genetic architecture of sleep health. A GWAS of the single-indicator factors would not add anything alone above and beyond the GWASs that were published by the original authors [9, 11, 12], but we include those single-trait GWASs (standard deviation of actigraphy sleep duration, self-reported sleep duration, and insomnia) in our post-GWAS analyses to compare biological pathways and external correlates across all domains of sleep health. By examining single nucleotide polymorphisms (SNPs) specific to the factors of circadian preference, alertness, and efficiency (as well as the single-indicator factors of regularity, non-insomnia, and duration) we assess, at a broader level, the molecular systems involved across sleep health domains as opposed to trait-specific pathways or genes; we also characterize whether these factor-level pathways and genes are consistent across the domains of sleep and circadian health.

Materials and Methods

Data

We utilized publicly available summary statistics from 12 previously published sleep GWASs that can be downloaded from https://sleep.hugeamp.org/downloads.html and are described in more detail in the referenced papers. Self-reported phenotypes were chronotype (no 23andme samples included), daytime sleepiness, sleep duration, napping, and insomnia [9, 10, 12, 13, 17]. Objectively measured sleep phenotypes estimated from actigraphy data were mid, L5, M10, efficiency, episodes, diurnal inactivity, sleep duration, and the standard deviation of actigraphy-based sleep duration [11]. M10, L5, mid, episodes, daytime sleepiness, napping, diurnal inactivity, insomnia, and the standard deviation of sleep duration were reverse-coded to reflect good health, as described in Morrison et al. [6]. Genetic correlations between all sleep indicators are depicted in Supplementary Figure S1.

Multivariate genome-wide association analyses

We used genomic SEM [16] in R (version 4.2.1) to conduct GWASs on the three multivariate sleep health factors from a genomic SEM (χ²(40) = 947.66, CFI = 0.950 SRMR = 0.071) constructed in Morrison et al. [6]. Data were organized using the sumstats() function to merge GWAS output files with the 1000 genomes phase 3 European reference panel and aggregate effect sizes and standard errors for each trait per SNP [16]. These traits were then munged with the munge() function that creates standardized “.sumstats” files for each trait containing SNP, z-scores and p-values. Then, the ldsc() function was used to create a covariance structure that contains the genetic covariance matrix and a sampling covariance matrix, used to account for any potential sample overlap, for all above-specified traits. Finally, we used the userGWAS() function in Genomic SEM to specify a model using the created covariance matrix then regress factors on all of the SNPs from the created sumstats file. We specified our latent structure of sleep health [6] and regressed the three multivariate factors—circadian preference, alertness, and efficiency—on 7,411,069 SNPs.

We computed effective sample size (N_effective) for each multivariate factor to be used as input for post-GWAS analyses using the equation from [15]. This formula computes effective N as the mean of $\frac{1}{2pq\ast S{E}^2}$ per SNP, after subsetting to include only SNPs with minor allele frequencies between 0.1 and 0.4.

Tests of heterogeneity

An advantage of using Genomic SEM is the ability to specify multiple genetic factor structures. We specified an independent pathways model where each SNP predicted all 12 indicators rather than the factors to test the assumption that the latent factors mediate the effects of the SNPs on the indicators (i.e. to test for heterogeneity of SNP effects); if a SNP is not acting through the factor, but instead is related to the indicators within the factor in a pattern that is inconsistent with their factor loadings (e.g. is only related to one indicator, or is related to multiple indicators but disproportionately to their loadings on the factor), it is said to be heterogeneous, sometimes referred to as a “Q SNP.” Q SNPs were identified here by first computing 6-df χ² difference tests (Δχ²) for each SNP, where in the more restrictive model the SNP predicts each factor and in the less restrictive model the SNP predicts each indicator directly instead of the factors. A significant Δχ² suggests that SNP does not act through the factors.

We ran follow-up analyses for significant Q SNPs that also significantly predicted one of the multivariate Genomic SEM factors. For each SNP, we regressed all factors on the SNP except the factor that had a significant hit for that SNP in its multivariate GWAS; for the latter factor, we regressed its indicators on the SNP instead of the factor. This more precise heterogeneity test tests whether the SNP is heterogeneous specifically for the factor it predicts.

Finally, we compared the “observed” versus “expected” SNP effects of the significant Q SNPs that also predicted one of the multivariate Genomic SEM factors. For each Q SNP and each indicator, we calculated the SNP effect expected by the factor model (the beta of the SNP predicting the factor multiplied by the factor loading for that indicator) and compared it to the SNP effect observed in the model in which the SNP directly predicted the indicators instead of the factors (the beta of the SNP predicting that indicator). If the expected beta is greater than the upper bound of the 95% confidence interval for the observed beta, the observed SNP effect is being under-estimated. If the expected beta is less than the lower bound of the 95% confidence interval for the observed beta, it is being overestimated.

Functional and biological characterization of genetic variants

We removed significant Q SNPs, and SNPs in LD (r² > 0.6) with those Q SNPs [18] from the output of the multivariate GWASs and ran the summary statistics through the standard SNP2GENE FUMA pipeline using the 1000 genomes phase 3 European reference panel. This method identifies genomic loci, lead SNPs, and independent significant SNPs associated with the submitted trait. Additionally, we ran the publicly available summary statistics from the original insomnia, sleep duration, and standard deviation of sleep duration GWASs through this same FUMA pipeline to compare results across factors. However, we reversed the effect sizes of the summary statistics for insomnia and the standard deviation of sleep duration so correlations would be in the health direction. This does not change the outcome of the analyses but was helpful for comparing results across domains. The parameters used for SNP2GENE analyses were r² < 0.6 to define independent significant SNPs, r² < 0.1 to define lead SNPs, and distance between LD blocks > 250kb.

The GENE2FUNC MAGMA pipeline is an extension of the SNP2GENE pipeline. It was employed to compute gene association tests, compute gene set enrichment, and determine tissue specificity in both our multivariate GWASs (after Q SNPs and SNPs in LD with Q SNPs had been removed) and the original (but reverse coded) insomnia, self-report sleep duration, and reverse-coded standard deviation of sleep duration GWASs. We specified GENE2FUNC (use all protein-coding genes from Ensembl version 92 and excluding the major histone complex region of chromosome 6) to annotate genes based on the PsychENCODE dataset and perform expression quantitative trait locus (eQTL) mapping for GTEx version 8 brain tissues (nominal p-value cutoff < 1e-3). We assessed significant gene sets based on the three Gene Ontology (GO) processes using MAGMA’s significance threshold of 0.05 for the corrected p-value. The GO processes are molecular (activities such as catalysis or transport that occur at the molecular level), cellular (representing locations relative to a cell where genes perform their functions), and biological (often larger and composed of multiple molecular processes) [19, 20].

Transcriptome-wide association studies are used to integrate expression quantitative trait loci (eQTL) data with GWAS signals [21]. MultiXcan is a program that tests how gene expression across tissues might play a mediating role in complex traits [22]. It uses trained linear prediction of gene expression from tissue-based data such as Genotype-Tissue Expression project (GTEx) and correlates those associations with the trait of interest [22]. We ran MultiXcan on all 6 sleep health domain summary statistics via the Complex Trait Genetics virtual lab (CTG-VL; https://vl.genoma.io/) to assess their expression levels in 13 brain tissues. Because there are numerous genes expressed in each tissue, we used the standard genome-wide alpha (p < 5e-8) for significance.

Classification of variants not found in prior sleep health analyses

To further classify how the multivariate GWASs might be similar or divergent from the already published sleep trait GWASs, we performed a Phenome-Wide Association Study (PheWAS) on genomic loci that were significant in our multivariate GWASs but not significant in the component indicator GWASs. We used MRBase’s “ieugwasr” R package to perform this analysis [23]. This package reads in a list of SNPs, in this case, loci that reached significance in the Genomic SEM factor GWASs but not the component GWASs, and uses an API token to pull any publicly available phenotypes (N phenotypes = 1144) that are also significantly associated with those loci [23].

Linkage disequilibrium score regression

Finally, to characterize the multivariate factors, we performed batch Linkage disequilibrium score regression (LDSC) on the multivariate GWAS summary statistics. LDSC, the method that Genomic SEM draws upon, computes heritability and bivariate genetic correlations between traits [24]. It takes summary statistics and uses a reference panel to look for patterns of associations across SNP effect sizes between two traits. We employed batch LDSC from the CTG-VL server (https://vl.genoma.io/) to estimate genetic correlations between sleep health and a host of behavioral and physical traits. With this batch approach, we computed genetic correlations between our sleep health factors and 181 external traits (significance determined by Bonferroni correction of 0.05/(181*3)). We uploaded summary statistics from the genomic SEM GWAS for circadian preference, alertness, and efficiency after removing Q SNPs and SNPs in LD with Q SNPs, and summary statistics from the original sleep duration [9], insomnia [12], and standard deviation of sleep duration [11]. Following Morrison et al. [6], we reverse-coded the direction of the effects in the summary statistics for insomnia and standard deviation of sleep duration so higher scores would reflect better sleep health (non-insomnia and regularity, respectively).

Results

Multivariate genome-wide association analyses

We computed multivariate GWASs on three factors of sleep and circadian health: circadian preference, alertness, and efficiency. The Circadian Preference factor (N_effective = 79 508) had 98 significant genomic risk loci. Of those 98 loci, 30 were significant loci in the constituent chronotype GWAS, 2 were significant loci in the constituent L5 GWAS, and 0 were significant loci in either the M10 or midpoint constituent GWASs (Figure 2). Sixty-seven significant loci associated with this factor were not significant in any of the constituent indicator GWASs. On the other hand, there were 161, 4, 1, and 1 loci that reached genome-wide significance in the constituent chronotype, least active 5 hours, most active 10 hours, and sleep midpoint GWASs, respectively, that did not reach significance in the circadian preference Factor GWAS.

Figure 2.

Manhattan plots and number of significant genomic loci for each multivariate genomic structural equation model (SEM) genome-wide association studies (GWAS) as well as the component GWASs. Manhattan plot depicts −log 10 p-values for each single nucleotide polymorphism (SNP) in the GWAS organized by chromosome. Shown in standard Manhattan Plot style are the loci for the multivariate Genomic SEM GWASs. The various overlayed dots represent loci that reached genome-wide significance in the component GWASs.

We found only eight significant genomic loci in the efficiency factor GWAS (N_effective = 31 654). Of those, one was a significant locus in the efficiency constituent GWAS and three were significant loci in the episodes constituent GWASs. Of the eight significant genomic risk loci, four were novel, or not significant in either of the indicator GWASs (Figure 2). There were 4 and 19 genome-wide significant loci in the efficiency and episodes constituent GWASs that did not reach genome-wide significance in the efficiency factor GWAS.

Finally, we found 85 significant genomic risk loci in the alertness factor GWAS (N_effective = 1,043,563). Of those, 8 were significant loci in the daytime sleepiness constituent GWAS, 31 were significant loci in the napping GWAS, and 0 were significant loci in the diurnal inactivity constituent GWAS. Of the 85 significant loci, 47 were novel, or not significant in any of the constituent indicator GWASs (Figure 2). Contrastingly, there were 29, 2, and 95 genome-wide significant loci in the alertness, diurnal, and napping GWASs that were not significant in the alertness factor GWAS. Overall, these findings indicate the multivariate factors were capturing genomic signal proportional to the amount of genomic signal in the indicators, but also were leveraging the boost in power to detect signal specific to the sleep and circadian domains that were not discovered before at the univariate level.

Tests of heterogeneity

We found 1322 significant Q SNPs (all p < 5e–8). However, only three of these Q SNPs were also independent significant SNPs for any of the three multivariate factor GWASs run here. Circadian preference was significantly predicted by two Q SNPs (rs12654450 & rs182588061); alertness was significantly predicted by one Q SNP (rs72820274); and efficiency was significantly predicted by one Q SNP (rs182588061), which also significantly predicted circadian preference. These results indicate that most of the genetic variants for these factors appear to be associated with the factors (i.e. they are associated with the indicators in patterns that are consistent with their loadings on the factors).

For those three significant Q SNPs, we ran follow-up nested model comparisons in which the Q SNP for a factor predicted the indicators for that factor instead of the factor, while also predicting other sleep health factors for which it was not significant. For example, SNP rs182588061was allowed to predict all factors except circadian preference (alertness, efficiency, duration, non-insomnia, and regularity), but the indicators of circadian preference (M10, L5, chronotype, and sleep midpoint). Figure 3 shows the observed SNP effects (betas referenced above) plotted with the SNP effects expected by the factor (gray dots; computed by multiplying the indicator loading by the SNP effect on the factor for each SNP and indicator) layered on top.

Figure 3.

Observed versus expected Q single nucleotide polymorphism (SNP) effects for the multivariate genomic structural equation model (SEM) genome-wide association studies (GWAS) factors. A Q SNP is one that is said to be heterogeneous, or not acting through the factor. Each bar graph represents the beta weight (and 95% confidence interval) of the effect of the particular SNP on the indicator from the Q SNP model. The Q SNP model is one where a SNP predicts all factors except the one it is presumed to be Q for, and instead predicts all indicators on that factor. The gray dots represent the expected betas for each indicator if the SNP acts through the factor (the beta of the SNP predicting the factor multiplied by the factor loading for that indicator).

The Q SNP chi-square difference test of the circadian preference factor was significant for SNP rs182588061, Δχ(3) = 76.38, p = 1.83e–16. SNP rs182588061 predicted L5 with a larger beta (β = 0.086, p = 1.35e–19) than it predicted chronotype (β = 0.015, p = .016), M10 (β = 0.021, p = .051), or midpoint (β = 0.020, p = .064). As shown in Figure 3, this pattern is inconsistent with the effects predicted by the loadings; because these indicators showed similar factor loadings on the circadian preference factor, the model predicts similar betas for all four indicators. The Q SNP chi-square difference test was not significant for SNP rs12654450, Δχ(3) = 2.27, p = .517, suggesting that it predicted the indicators—chronotype (β = −0.013, p < .001), midpoint (β = −0.010, p = .001), L5 (β = −0.006, p = .039), and M10 (β = −0.005, p = .060)—in proportion to their loadings on the circadian preference factor.

The Q SNP chi-square difference test of the alertness factor was also significant: Δχ(2) = 45.27, p = 1.48e–10. SNP rs72820274 predicted daytime sleepiness (β = 0.004, p = 1.72e–12) with a nominally larger beta than it predicted napping (β = 0.002, p = .001) and diurnal inactivity (β = 0.001, p = .678), whereas the factor loadings would predict more similar associations, as shown in Figure 3. SNP rs182588061 was not a significant Q SNP for the efficiency factor: Δχ(1) = 1.69, p = .193, suggesting that it predicted the indicators in proportion to their loadings on the efficiency factor: efficiency (β = 0.079, p = 9.21e–17), and episodes (β = 0.031, p = .002). regardless of whether the Q SNP follow-up tests were significant, we excluded all three of these SNPs from follow-up analyses of the factors they significantly predicted based on the fact that they met criteria as a Q SNP at the full 6-df Q SNP test level.

Overall, the relative dearth of significant Q SNPs indicates the multivariate factors are indeed capturing genomic signals specific to the sleep and circadian constructs being extracted. If heterogeneity were widespread, multivariate factors would likely not be useful or adding information above and beyond the sum of the individual indicator GWASs.

Functional characterization of genetic variants and genes

To characterize the multivariate sleep health factors, we ran our summary statistics through the FUMA/MAGMA SNP2GENE and GENE2FUNC pipelines, after excluding significant Q SNPs (m = 1322) and any SNPs that were in LD (r² > 0.6) with those Q SNPs (total SNPs excluded m = 3189). We also ran the single-indicator summary statistics from the original indicator GWASs through the same pipelines to compare functional characterization of all domains of sleep health. We did not find any significant genomic signal associated with regularity, consistent with the original GWAS of the standard deviation of sleep duration [11], so we do not include it for the FUMA/MAGMA results.

We found that 311 SNPs were significant genomic risk loci in at least one factor GWAS. Of those, only three were genome-wide significant loci for more than one factor. SNP rs12140153 was associated with both the alertness and circadian preference GWASs, and rs7556815 and rs75606464 were associated with non-insomnia and duration GWASs.

Similarly, 3391 genes were significantly enriched for at least one sleep health factor. Of those, 75% (N_genes = 2539) were only significantly enriched for one sleep health factor. There were eight genes associated with four factors, 147 genes associated with three factors, and 697 genes associated with two factors (Figure 4). The eight genes associated with four factors were AKTIP, HOOK1, IRX5, IRX6, KIAA1841, MMP2, PAPOLG, and RBL2, and they were all significant in the circadian preference, alertness, efficiency, and duration factors.

Figure 4.

Upset plot of overlapping significant genes between sleep health factors. The bars on top represent how many genes were significant for each factor or combination of factors. The lines connect factors (filled in dots) that have shared significant genes.

Despite few shared significant loci between sleep and circadian preference domains, we did see some overlap at the gene, gene set, and enrichment levels. These results suggest that perhaps the top significant SNPs associations with sleep health domains are not all the same, but there are still common patterns of associations between SNPs and genes contributing to the formation of sleep health behaviors. This suggestion is consistent with the weak to moderate genetic correlations between factors (|rGs = 0.04–0.51|).

We assessed the biological, molecular, and cellular GO pathways enriched across sleep health and circadian domains and found some overlap, but mostly heterogeneity. The most overlap came in the GO biological process, which are the broadest gene set pathways. Circadian preference, alertness, efficiency, and non-insomnia were all associated with telencephalon development, and circadian preference and efficiency were both associated with rhombomere development, both biological GO sets involved in brain structure maturation. Efficiency was associated with the gene-silencing biological set, while circadian preference and duration were associated with epigenetic regulation of gene expression. Only Alertness was associated with protein tetramerization and regulation of megakaryocyte differentiation, and only duration was associated with DNA conformation change and chromosome organization. Efficiency and duration were significantly associated with chromatin organization and regulation of gene silencing.

Circadian preference showed enrichment in several sleep and circadian-related biological GO sets: entrainment of circadian clock by photoperiod, entrainment of circadian clock, positive regulation of circadian rhythm, negative regulation of circadian rhythm, circadian rhythm, regulation of circadian rhythm, circadian sleep–wake cycle sleep, circadian sleep–wake cycle, and regulation of circadian sleep–wake cycle. Non-insomnia had genes enriched in six of those pathways, duration had genes enriched in two of those pathways, alertness had genes enriched in four of those pathways, and efficiency had genes enriched in one of those pathways. In addition to these shared pathways with circadian preference, non-insomnia was enriched in the circadian sleep–wake cycle REM sleep pathway, and alertness was enriched in the circadian sleep–wake cycle non-REM sleep and negative regulation of circadian sleep–wake cycle sleep pathways. The overlap of enrichment between sleep and circadian health domains demonstrates those groups of domains are not entirely distinct and do share molecular makeup. Furthermore, the finding that alertness and non-insomnia were enriched in circadian-related pathways in which circadian preference was not enriched shows sleep health is intrinsically tied to circadian health.

Within molecular GO terms, circadian preference genes were significantly enriched in the zinc ion binding and transition metal ion binding gene sets. Duration genes were significantly enriched in the protein heterodimerization activity and mRNA binding sets. Overall, across all factors, there was divergence in the GO set enrichment patterns. Biological terms saw a handful of overlap between factors, and genes for each factor were significantly enriched in at least one biological set. Cellular terms saw more similar patterns for alertness and duration than other factors, and molecular terms only saw significant gene enrichment for circadian preference and duration.

Lastly, within the GO cellular terms, genes related to alertness, efficiency, and duration perform their functions in the nuclear chromosome, whereas genes associated with just alertness and efficiency perform their functions in the chromatin. Finally, only genes associated with Duration significantly perform their functions in the chromosome. Figure 5 plots the −log10 p-value for any pathway that was significant (MAGMA uses a significance threshold of 0.05 for the corrected p-value).

Figure 5.

Gene Ontology (GO) terms significantly enriched across sleep health factors from functional mapping and annotation of genome-wide association studies (FUMA/MAGMA). Dotted line is the corrected significance threshold of p = .05 (on the −log 10 scale). Go terms are categorized by biological, cellular, and molecular sets.

Finally, we performed a Transcriptome-Wide Association Study (TWAS) using MultiXcan [22] to examine how gene expression across 13 brain tissues might mediate the relationships with sleep health. The cerebellum had the greatest number of significant genes expressed for all sleep health domains except efficiency (N_genes = 11–28). Regularity did not register any significant genes expressed in any of the 13 brain tissues. The gene with the lowest p-value for alertness was SPPL2C, expressed in the frontal cortex; the gene with the lowest p-value for circadian preference was RP11-624D11.2, expressed in the basal ganglia; the gene with the lowest p-value for duration was EIF3KP1, expressed in the cerebellum; the gene with the lowest p-value for efficiency was GNL3, expressed in the cerebellar hemisphere; and the gene with the lowest p-value for Non-Insomnia was NAT6, expressed in the amygdala. Figure 6 plots the TWAS as a Manhattan plot with brain regions organized on the x axis, alongside bar plots showing the number of significantly enriched genes per region.

Figure 6.

Transcriptome-Wide Association Study (TWAS) Manhattan plot (all genes enriched per tissue type on x axis, −log 10 p-values on y axis), and bar graphs depicting number of genome-wide significant genes enriched per tissue type.

Classification of variants not found in prior sleep health analyses

A major motivation for this project was to assess how leveraging genomic SEM and a multivariate GWAS might inform in a similar or different manner from the component univariate sleep health GWASs. To do so, we performed a PheWAS on the significant sleep health loci for factors that did not reach genome-wide significance in the respective constituent GWASs. As previously mentioned, circadian preference had the highest number of novel loci, followed by alertness and then efficiency. Correspondingly, PheWAS results showed the novel loci associated with the circadian preference factor were associated with the highest number of external traits (2550 loci on 801 traits), followed by alertness (1705 loci on 557 traits), and efficiency (149 loci on 99 traits). We hand-categorized these external traits into 17 categories based on our judgments, plotted in Figure 7. We created a category called “unclassified” that contained traits with uninformative labels, such as “ENSG00000179029” or “IGF-1.” The full list of traits, including those in the unclassified category, are shown in Supplementary Tables S1–S3.

Figure 7.

Phenome-wide association study (PheWAS) computed on novel loci for each genomic SEM sleep health factor using Mendelian-Randomization Base (MRBase). Trait categories were classified by hand and plotted as a Manhattan plot. All depicted loci-trait associations reached genome-wide significance.

The category with the most SNP hits across all three factors was anthropometric traits, followed by blood-assay traits, general health traits, and, not surprisingly, sleep-related traits. Although anthropometric was the category with the most loci-associated traits for all three factors, circadian preference seemed disproportionately related to anthropometric traits compared to alertness and efficiency. The association between circadian preference and anthropometric traits appears to be driven by BMI, and tops hits for circadian preference in the health category were also BMI-related such as obesity and diabetes. Cardiovascular traits, such as basal metabolic rate, heart rate recovery, and blood clot, showed a large peak for circadian preference as well, indicating the importance of the role of the sleep–wake cycle on heart health, which might, in turn, affect diabetes, obesity, and BMI [25–27]. Top hits in the sleep category for circadian preference were all morning–evening preference and napping during the day.

Novel alertness loci were also significantly associated with anthropometric and blood-assay-related traits, but the individual traits within those categories differed from those related to circadian preference: alertness loci were associated with lung function/oxygen traits, such as hematocrit, red blood cell concentration, and forced expiratory volume. Within the anthropometric category, the top hits for alertness were also related to body composition but were slightly difference from BMI (measures of impedance). The alertness loci were also associated with sleep traits such as napping during the day.

The efficiency loci were associated with much fewer traits, due to the smaller number of novel loci, but still showed links with anthropometric and cardiovascular traits. However, efficiency was more related to impedance, like alertness, and cholesterol levels than to BMI or heart/lung function directly. Other top hits for the novel efficiency loci were osteoarthritis and interleukin-1 receptors, which both relate to inflammation response [28]. In the sleep category, efficiency loci were related to only measures of sleeplessness/insomnia.

We also performed ad hoc PheWASs on loci that reached genome-wide significance for the indicators of a factor but did not reach significance for the factor. The results of these subsequent PheWASs largely mirror those of the novel loci PheWASs. The indicator PheWASs show more trait associations (due to the larger number of significant loci) but reflect similar patterns to the novel loci PheWAS (Supplementary Figure S5). Of all the traits associated with each novel-loci PheWAS, 88%–90% were also associated with loci found to be significant in the indicators but not that factor. These indicator loci appear to be capturing similar genomic associations to the factors, but also might represent genetic signals more specific to an indicator because it did not reach genome-wide significance in the factor. We also performed a PheWAS on the original insomnia, self-report, and actigraphy-based duration GWASs, which are presented in Supplementary Materials and Tables S4–S8.

Linkage disequilibrium score regression

Finally, to classify external correlates across sleep health domains, we used linkage disequilibrium score regression (LDSC) via the CTG-VL to compute genetic correlations between circadian preference, alertness, efficiency, duration, non-insomnia, and regularity with 181 external traits. This method runs batch genetic correlations for 1403 external traits (mostly from the UK Biobank), but we selected 181 a priori relevant traits to focus on. These traits were a mix of health diagnoses, physical and behavioral characteristics, environmental factors, and biochemical blood assays. Using a Bonferroni corrected alpha of 0.05/(181*6) = 4.604e-05, we found circadian preference was significantly associated with 17 traits, efficiency was significantly associated with seven traits, alertness was significantly associated with 55 traits, duration was significantly associated with 23 traits, regularity was significantly associated with 18 traits, and non-insomnia was significantly associated with 68 traits.

Sleep and circadian health were overwhelmingly associated with psychiatric disorders and general health in a consistent manner. There were 23 psychiatric disorders and 12 general health problems associated with at least one sleep or circadian health factor. Furthermore, only 2 psychiatric disorders (0 general health problems) were positively correlated with sleep health: duration had positive correlations with schizophrenia and bipolar disorder.

Interestingly, 64% (N_phenotypes = 50) of the 78 phenotypes significantly associated with at least one factor were significantly associated with another factor. There were no phenotypes significantly associated with all six factors, but BMI, educational attainment, and Townsend deprivation index at recruitment were significantly associated with five factors (BMI: all except circadian preference; educational attainment and Townsend deprivation index: all except efficiency). The lack of genetic correlations between circadian preference and BMI is surprising, given that circadian preference novel loci appeared to be robustly associated with BMI in the PheWAS. Associations between sleep health and BMI and Townsend deprivation index were all negative, but interestingly, educational attainment was negatively correlated with circadian preference (lower educational attainment genetic liability associated with liability for being a morning person), but was positively associated with duration, alertness, non-insomnia and regularity (higher educational attainment genetic liability associated with better sleep health genetic liability). Figure 8 shows genetic correlations (and 95% confidence intervals) for any phenotype that was significantly associated with at least one factor but does not show the genetic correlations for any nonsignificant factor-phenotype associations for clarity.

Figure 8.

Linkage disequilibrium score regression (LDSC)-derived genetic correlations between sleep health factors and health and behavioral traits. Points are genetic correlations (rGs) and error bars are 95% confidence intervals. Traits are only shown if they were associated with at least one sleep health factor.

See Supplementary Tables S9–S13 and Figures S8–S10 for full list and plots of LDSC associations. From the full list of 1403 phenotypes, at least one sleep health factor was associated with 645 traits (at a new Bonferroni correction level of 0.05/(1,403*6)). Circadian preference was associated with 36, alertness with 187, efficiency with 19, duration with 64, non-insomnia with 333, and regularity with 6 traits.

Relationships to actigraphy duration

Self-reported and actigraphy-based sleep duration do not correlate strongly with each other or external traits [29]. Using available GWAS data, the genetic correlation is significant, but not large (rG = 0.43), suggesting that objective and subjective measures might capture somewhat distinct constructs. We tested multiple model structures in Morrison et al. [6] in an attempt to include as many sleep health traits as possible, while still capturing meaningful factors. Ultimately, we decided against including actigraphy sleep duration because other indicators in the model (efficiency and episodes) were derived from actigraphy sleep duration, leading to idiosyncratic patterns of correlations (both positive and negative) with these indicators that caused model fit problems. Although the final model in Morrison et al. [6], which is the basis for our factor GWASs, did not include actigraphy duration, in this section we compare the post-GWAS analyses to the same analyses for actigraphy duration.

Results are presented in depth in Supplementary Materials, but briefly, actigraphy duration appeared to show largely divergent genetic architecture compared to not only the other sleep and circadian health factors, but also self-reported duration. Looking across all sleep and circadian domains (now including actigraphy-based duration), there were 320 SNPs associated with any domain, and only five SNPs associated with two domains. Actigraphy duration had one SNP in common with efficiency and one SNP in common with non-insomnia. Actigraphy duration had 31 genes that were also significantly associated with self-reported duration, but only 10 of those were unique to the two duration constructs. There were 23 genes shared among actigraphy duration, self-reported duration, and efficiency, which makes sense given sleep efficiency and sleep episodes were phenotypically derived from the actigraphy duration variable. However, the gene set analyses for GO terms, TWAS of brain regions, batch LDSC analysis, and PheWAS showed largely different results between actigraphy and self-reported duration (Supplementary Materials), indicating divergence at the molecular level between actigraphy duration and other sleep and circadian health domains. These analyses provide valuable information as to how the genetic architecture of sleep duration varies based on the measurement method used.

Discussion

We performed three multivariate GWASs of sleep and circadian health factors (circadian preference, alertness, and efficiency) using genomic SEM. Bioinformatic analyses on the multivariate factors and three single-indicator traits (non-insomnia, duration, and regularity) revealed the most genomic overlap across these domains at broader levels—such as significant enrichment of the cerebellum and basal ganglia and negative genetic correlations with psychiatric disorders across all sleep health domains—but less overlap at the actual level of the gene or gene-sets. Generally, all sleep and circadian health domains were related to anthropometric, cardiovascular, and health traits, but circadian preference showed the most associations with obesity-related cardiovascular traits (e.g. BMI, obesity, and heart rate), while alertness related more to lung-function traits (e.g. blood oxygen levels, forced vital capacity) and efficiency related more to inflammatory response traits (e.g. cytokines and arthritis). In previous work, we found the sleep health genomic factors studied here relate in a similar manner to psychiatric domains, but still predicted unique variance in psychiatric traits, especially when controlling for correlations among sleep health [6]. The results of this paper extend that work and demonstrate that generally the domains of sleep health share some analogous genetic makeup and relate comparably to external trait categories, but also diverge enough to warrant studying them as distinct but related entities.

No single sleep trait exists in a vacuum. For example, two people who have completely opposite circadian preferences may present alike or inversely in terms of their alertness, efficiency, or duration. Nonetheless, sleep is complex ecological process and the intrinsic domains of sleep health do show associations with each other, phenotypically, and genetically. Here, we characterize the similar but diverse genetic architecture of sleep health. The six domains we studied show support for general resemblance between sleep health domains but still highlight the distinct genomic variance between domains. For example, even though circadian preference, alertness, and efficiency novel loci were all associated with the cerebellum and anthropometric, cardiovascular, and health-related categories, the individual traits show interesting and divergent patterns of associations. Sleep and the circadian system have been related to a wide range of cardiovascular diseases and endocrine activity that evokes fluctuations in immune parameters such as T cells, proinflammatory cytokines, and more [30, 31], and can be corroborated here by the various PheWAS and LDSC associations across all domains. Identifying the unique associations between sleep health domains and other health conditions has the potential to inform public health interventions that treat sleep as a modifiable health behavior in order to reduce known health disparities [30].

Telencephalon development was the gene set that was most associated with all sleep and circadian health domains except duration. The telencephalon is a part of the brain that encompasses the cerebral cortex, hippocampus, and basal ganglia. This result is in line with prior research establishing that sleep regulation occurs in the cerebral cortex and is a restorative activity that supports brain health and cognition [32, 33]. Similarly, the cerebellum was by far the region with the most gene expression across all sleep health factors. The next several regions with significant gene expression were the cerebellar hemisphere, cortex, and both the caudate basal ganglia and nucleus accumbens basal ganglia. The cerebellum is not studied as frequently as cortical structures in conjunction with sleep health, yet it has been implicated in several sleep disorders including sleep apnea, chronic insomnia, fatal familial insomnia, and restless leg syndrome [34, 35]. These robust associations point towards the role of the cerebellum in not only sleep disorders, but the broader construct of sleep health too. The basal ganglia consist of several subcortical structures and have been implicated in the sleep–wake cycle in previous findings as well [36].

Based on several PheWASs, anthropometric traits were the category most associated with sleep and circadian health. However, anthropometric associations with circadian preference were mostly driven by BMI, which has been associated with sleep traits [37], whereas alertness and efficiency were most associated with measures of impedance (total body water) within the anthropometric category. Circadian preference was more associated with obesity-related traits in general. The most associated cardiovascular traits for circadian preference were basal metabolic rate and heart rate recovery, and the most associated traits in the health category were obesity and diabetes. In contrast, the top associated traits in the blood assay and cardiovascular categories for alertness were hematocrit, red blood cell counts, and lung function/capacity. On the other hand, efficiency was more associated with cholesterol levels, osteoarthritis, and interleukin inflammatory receptors, perhaps indicating the role of sleep efficiency in stress response [28].

Psychiatric and general health-related categories were overwhelmingly negatively related to sleep and circadian health, such that genetic liability for better sleep health reflects genetic liability for better physical and mental health. The two exceptions were that schizophrenia and bipolar disorder had positive correlations with duration, which has been found in other genetic studies [9]. These genetic associations do not necessarily imply better sleep is associated with psychiatric disorders, because too much sleep is often related to negative health outcomes [38].

Although for the most part psychiatric traits were negatively correlated with sleep and circadian health, several interesting associations emerged. non-insomnia, alertness, and duration were associated with most or all neuroticism score items (nervous feelings, mood swings, miserableness, loneliness, isolation, irritability, guilty feelings, and fed-up feelings) as well as the overall neuroticism score. Interestingly, regularity was not associated with the overall neuroticism score, but was correlated with irritability and mood swings—two neuroticism items that reflect mood variability. This pattern of associations demonstrates how the genetic relationships between sleep health and psychiatric disorders, while generally negative, are also nuanced and not always consistent across both sleep and psychiatric disorder domains.

The breadth of traits studied here can help inform how sleep health relates to a wide variety of outcomes that are perhaps not usually studied. Even several air pollution variables were found to have negative genetic correlations with sleep, such that genetic liability for sleep regularity, daytime alertness, duration, and non-insomnia are associated with lower liability to particulate matter (2.5) air pollution exposure. Genetic liabilities for non-insomnia and alertness were also positively genetically associated with exercise types and healthy social activities like sports clubs or adult education classes.

Taken together, this paper provides a comprehensive analysis of the genetic architecture of sleep health. By leveraging previously published sleep health trait GWASs and performing multivariate GWASs on novel sleep factors, we were able to characterize some of the biological functions and external correlates related to sleep health. The results suggest there are broad patterns of shared biological and molecular function across domains of sleep health, but also genomic differences across sleep health. For example, we did not see common significant gene sets across factors that also are associated with common correlates of sleep health, such as GABAergic processes and psychiatric disorders [39]. Instead, we saw a plethora of variegated associations without clear overlap throughout the post-GWAS analyses. Here, we corroborated previous findings of unique patterns of associations between specific sleep and psychiatric domains by recapitulating how the genetic makeup of sleep health differs across factors.

Limitations

Most limitations of this study are in the form of interpreting the genomic SEM results. Structural equation modeling is a method that assumes covariances among indicators arise because they are caused by unobserved traits. Performing a genomic SEM GWAS involves calculating genetic variance in a latent, or unobserved trait. Heritabilities are not calculated from genomic SEM GWASs because the total variance in the latent factor is only genetic. That is, because the indicators are GWAS summary statistics themselves, the factors do not capture environmental variance unless that environmental variance is correlated with the genetic influences. However, in a previous paper, we wrote about testing multiple genomic structures of sleep health in order to instill more confidence in the validity of the latent sleep health factors [6]. Furthermore, Q SNP analyses in this study revealed that sleep health factors were largely acting through the factors, and not disproportionately through specific indicators.

A major limitation of this study is that it only includes one conceptualization of sleep and circadian health. Because the sleep health domains encompass many diverse sleep traits, there are multiple permutations of observed variable combinations that could be used to capture sleep and circadian health. For example, in this model, we included self-reported sleep duration but excluded actigraphy-based sleep duration for reasons discussed in depth in Supplementary Materials, although we did incorporate comparisons with actigraphy-based sleep duration in the Results. Because there is no gold standard for parameterizing sleep health, there is no right or wrong combination of variables. This limitation is an important point to keep in mind when considering the interpretation of these sleep health factors and how they relate to external traits. However, one major benefit from the push for open science is the public availability of sleep and circadian health summary statistics. We encourage other research to replicate or extend this model by exploring how other sleep-related traits may fit alongside these factors.

Additionally, we are not able to assess causality in this study. The post-GWAS analyses allow us to characterize and compare the genetic makeup of sleep health domains and their relations to external traits. However, these analyses are correlational so we cannot conclude whether any traits cause liability for other traits. To assess causality other models, such as randomized control trials, co-twin control, or Mendelian Randomization, are needed.

Other limitations come from the nature of whole-genome data. Because all the data in this study was previously published, it is subject to any limitations or biases from the original samples themselves. For example, many GWASs were calculated in the UK Biobank, which is known to have ascertainment bias. Similarly, because many of these measures were collected as part of large consortia, the variables are not as deeply phenotyped as they would be in smaller sleep studies. Finally, this study was conducted with all European samples, so may not capture the genomic makeup and relationships among sleep health in other ancestral groups. As sample sizes for non-European GWASs of sleep measures become available, it will be a priority to evaluate these models in those samples.

Conclusions

Although sleep health is frequently studied and associated with a host of health outcomes, the underlying etiology and biological pathways involved in the manifestation of sleep patterns have not been fully explored. In this paper, we showed how sleep health domains appear to show unique genetic makeup, biological pathways, and gene expression patterns. The sleep health domains studied here correlate weakly to moderately at the genetic level (|rGs| = 0.04–0.51), and this is reflected in their shared genomic enrichment in brain regions such as the cerebellum and basal ganglia, and associations with external traits. Overall, better sleep health liability is genetically associated with better health and behavioral liability. However, given the divergence in the underlying genetic architecture of sleep health domains, it is important to understand the nuanced relationships of these sleep health domains with specific external traits. A better comprehension of how sleep health cooccurs with so many other health outcomes, whether that is from shared genetic variation, common environmental exposures, or patterns more consistent with causality, will aid in the treatment and prevention of many costly, harmful disorders and behaviors.

Funding

CLM received support from National Institutes of Health: MH016880. EAW received support from MH015442 and DA017637. NPF received support from DA046064, DA046413, DA051018, DA042742, MH117131, HD078532, and AG046938.

Disclosures Statement

Financial Disclosure: KPW reports research support/donated materials: DuPont Nutrition & Biosciences, Grain Processing Corporation, and Friesland Campina Innovation Center. Financial relationships: consulting with or without receiving fees and/or serving on the advisory boards for Circadian Therapeutics, LTD., Circadian Biotherapies, Inc., and the U.S. Army Medical Research and Materiel Command–Walter Reed Army Institute of Research. The other authors declare no competing interests. Nonfinancial Disclosure: The authors declare no other competing interests.