Main

Rather than being 'secondary', disordered sleep may have an important role in the etiology and maintenance of physical and mental health1,2. Heritability has been estimated at ∼40% for sleep duration4,6,7,8, 25–45% for insomnia9 and 17% for excessive daytime sleepiness9, but few genetic factors are known. A Mendelian mutation in BHLHE41 (encoding p.Pro385Arg) associated with short sleep duration has been identified and was confirmed in mouse models10. Genome-wide association studies (GWAS) for sleep duration have been reported11,12,13,14, but only an association at the PAX8 locus reached genome-wide significance and was confirmed across ancestry groups12. There are several reported loci associated with restless legs syndrome (RLS) and narcolepsy, but no robust genetic loci have been associated with insomnia symptoms or excessive daytime sleepiness15,16.

We and others have performed a GWAS for chronotype in the UK Biobank17,18 and a 23andMe participant sample19. To identify genetic variants that contribute to self-reported sleep duration, insomnia symptoms and excessive daytime sleepiness and link them with other conditions, we performed GWAS using phenotype measures in UK Biobank participants of European ancestry. Variation in sleep duration, insomnia symptoms and excessive daytime sleepiness was significantly associated with age, sex, principal components of ancestry, genotyping array, depression, psychiatric medication use, self-reported sleep apnea and BMI (Supplementary Table 1), as previously reported20,21,22,23. Together, age, sex and principal components explained 0.4%, 3.0% and 1.3% of variation in sleep duration, insomnia symptoms and excessive daytime sleepiness, respectively. Strong and significant pairwise phenotypic correlation was seen between the traits overall and within each sex, with limited correlation observed with chronotype (Fig. 1 and Supplementary Fig. 1).

Figure 1: Sleep traits are phenotypically and genetically correlated.
figure 1

(a) Phenotypic correlation between the reported sleep traits, using Spearman correlation (r). (b) Genetic correlation (rg) between the reported sleep traits, using LD score regression67. The color scale represents the strength of correlation. Chronotype ranges from extreme morning types to extreme evening types.

Full size image

We performed GWAS analyses of sleep duration, insomnia symptoms and excessive daytime sleepiness using linear/logistic regression adjusting for age, sex, ten principal components and genotyping array. We identified 9 genome-wide significant (P < 5 × 10−8) and 14 suggestive (P < 5 × 10−7 to P = 5 × 10−8) loci (Fig. 2, Table 1 and Supplementary Figs. 2 and 3). For sleep duration (n = 111,975), the strongest association was observed at the PAX8 locus (rs62158211[T]: β (s.e.) = 2.34 (0.30) min/allele, P = 4.7 × 10−14, effect allele frequency (EAF) = 0.213; Fig. 2a), confirming a previously reported association (r2 = 0.96, D′ = 1 to lead SNP rs1823125 in 1000 Genomes Project CEU (European-ancestry) individuals)12. For insomnia symptoms (n = 32,155 cases and 26,973 controls), we observed significant associations within MEIS1 (rs113851554[T]: odds ratio (OR) (95% confidence interval (CI)) = 1.26 (1.20–1.33), P = 9.1 × 10−19, EAF = 0.057; Fig. 2b), a homeobox gene implicated in motor neuron connectivity in Drosophila melanogaster24,25, retinal and lens development in mouse26 and substance P expression in the amygdala in humans27; near TMEM132E (rs145258459[C]: OR (95% CI) = 1.23 (1.13–1.35), P = 2.1 × 10−8, EAF = 0.983; Fig. 2c), which belongs to a gene family whose members have roles in brain development28, panic/anxiety29 and bipolar disorder30, suggesting a link between insomnia symptoms and an underlying broader sensitivity to anxiety and stress; and near CYCL1 (rs5922858[G]: OR (95% CI) = 1.12 (1.07–1.16), P = 1.28 × 10−8, EAF = 0.849; Fig. 2d), a locus previously associated (P = 1 × 10−6) with alcohol dependence comorbid with depressive symptoms31. Sex-stratified analyses identified an additional female-specific signal near TGFBI (rs3792900[C]: OR (95% CI) = 1.10 (1.07–1.14), P = 2.16 × 10−8, EAF = 0.470; Table 1, Supplementary Fig. 3q,r and Supplementary Table 2), which encodes an extracellular matrix protein responsible for human corneal dystrophy32, and a male-specific signal near WDR27 (rs13192566[G]: OR (95% CI) = 1.14 (1.09–1.20), P = 3.2 × 10−8, EAF = 0.860; Table 1, Supplementary Figs. 3s,t and 4, and Supplementary Table 2), which encodes a scaffold protein. Independent associations at both loci are observed with type 1 diabetes, suggesting that these proteins may have a role in autoimmunity33,34,35. For excessive daytime sleepiness (n = 111,648), we identified a signal near the androgen receptor gene AR (rs73536079[T]: β (s.e.) = 0.634 (0.115), P = 3.94 × 10−8, EAF = 0.002; Fig. 2e), with no sex-specific effects. Secondary analyses after additional adjustment for depression or BMI identified a signal near ROBO1 (depression adjustment n = 107,440; rs182765975[T]: β (s.e.) = 0.099 (0.018), P = 3.33 × 10−8, EAF = 0.003; Table 1 and Supplementary Fig. 3o), which encodes a neuronal axon guidance receptor previously implicated in dyslexia36, and a signal near another member of the TMEM132 family, TMEM132B (BMI adjustment n = 75,480; rs142261172[A]: β (s.e.) = 0.106 (0.018), P = 9.06 × 10−9, EAF = 0.004; Table 1 and Supplementary Fig. 3p). Conditional analyses did not identify independent association signals (Supplementary Table 3). Sensitivity analyses adjusting for factors influencing sleep traits, including self-reported sleep apnea, depression, psychiatric medication use, smoking, socioeconomic status, employment status, marital status and snoring, did not significantly alter results for primary association signals (Supplementary Table 4).

Figure 2: Regional association plots for genome-wide significant loci.
figure 2

(a) Sleep duration. (b–d) Insomnia symptoms. (e) Excessive daytime sleepiness. (f,g) Composite trait of sleep duration, insomnia symptoms, excessive daytime sleepiness and chronotype. Chromosomal position is indicated on the x axis, and the −log10 P value for each SNP (filled circles or squares) is indicated on the y axis, with the lead SNP shown in purple (the 400-kb window around each lead SNP is shown). The genes within each region are shown in the lower panel. The blue lines represent the recombination rate. Additional SNPs in each locus are colored according to linkage disequilibrium (r2) with the lead SNP (estimated by LocusZoom on the basis of HapMap CEU haplotypes or within UK Biobank (c)). Squares represent directly genotyped SNPs, and circles represent imputed SNPs.

Full size image
Table 1 Genome-wide significant (P < 5 × 10−8) and suggestive (P < 5 × 10−7) loci associated with sleep duration, insomnia symptoms and excessive daytime sleepiness in subjects of European ancestry in the UK Biobank
Full size table

The leading associations overlap interesting candidate genes whose expression is enriched in hypocretin-expressing neurons in mice and zebrafish37,38, that are differentially expressed in sleep-deprived rats39 and/or that regulate sleep in Drosophila40. Credible set analyses41 highlighted a number of potential causal variants at each locus (Table 1), and future experimental studies will be necessary to characterize the functions of these variants in regulating sleep traits. Bioinformatic annotations42 offer an initial opportunity for in silico functional interpretation of the variants (Supplementary Fig. 5 and Supplementary Table 5). For example, multiple variants for all three traits are predicted to disrupt binding of FOXP1, a neural transcriptional repressor implicated in intellectual disability, autism and language impairment43. Interestingly, the locus associated with sleep duration encompassing PAX8 is adjacent to the only chromosomal fusion site that arose since the divergence of humans from other hominids ∼5 million years ago44,45, and the new genomic structure created by this unique evolutionary history may have a causal role in sleep duration. Pathway analysis46 of significant and suggestive loci identified enrichment of genes associated with immune, neurodevelopmental, pituitary and communication disorders (P < 0.01) and genes enriched for binding sites for the stress-responsive transcription factor heat shock factor 1 (HSF1) and the endoplasmic reticulum stress- and unfolded-protein-responsive transcription factor HERPUD1 (Supplementary Tables 6 and 7).

Aside from the lead SNPs in the PAX8 region and a variant in the DRD2 region47 for sleep duration, we found limited evidence of association for previously published candidate gene or GWAS signals (Pmeta < 5 × 10−5; Supplementary Table 8) or for regions encompassing core clock genes (Supplementary Fig. 6). Our findings for the GWAS on sleep duration largely overlap with those of Jones et al.18, despite differences in exclusion criteria and analytical approach. In particular, our study excluded shift workers (n = 6,557), sleep medication users (n = 1,184) and first- to third-degree relatives (n = 7,980), whereas the linear mixed-model analyses by Jones et al.18 included these populations, leading to a larger sample size (n = 127,573). Likely because of this increased power, Jones et al.18 identified two additional signals at VRK2 that did not attain genome-wide significance in our study (rs1380703[A]: β (s.e.) = 1.50 (0.30) min/allele, P = 8.43 × 10−8 and rs17190618[T]: β (s.e.) = 1.60 (0.34) min/allele, P = 3.80 × 10−6).

Trait heritability calculated as the proportion of trait variance due to additive genetic factors measured here (observed-scale SNP heritability, h2 (s.e.)) was 10.3% (0.006%) for sleep duration, 20.6% (0.011%) for insomnia symptoms and 8.4% (0.006%) for daytime sleepiness (BOLT-REML variance-components analysis48). LD score regression analysis49 confirmed no residual population stratification (intercept (s.e.): sleep duration, 1.012 (0.008); insomnia symptoms, 1.003 (0.008); excessive daytime sleepiness, 1.005 (0.007)). Tests for enrichment of heritability by functional class using an LD score regression approach50 identified excess heritability across active transcriptional regions for insomnia symptoms and across genomic regions conserved in mammals for all three sleep traits. Consistent with these findings, heritability enrichment in conserved regions was seen for traits demonstrating significant genetic correlation with sleep (Fig. 3 and Supplementary Table 9).

Figure 3: Partitioning of the genetic architecture of sleep duration, insomnia symptoms and excessive daytime sleepiness across functional annotation categories.
figure 3

For each trait, fold enrichment estimates for the main annotations from LD score regression50 are indicated on the y axis across functional annotation classes on the x axis. Error bars represent the 95% confidence intervals around the estimates. We tested 25 functional annotations; annotations passing the multiple-testing significance threshold (P < 0.005) are shown. For context, the average enrichment across functional annotation categories is shown for nine traits with significant genetic correlation to at least one sleep trait (GWAS traits correlated with sleep; including GWAS for BMI, waist circumference, birth weight, depression, educational attainment, three glycemic traits in non-diabetics and schizophrenia) and for five traits with no significant genetic correlation to any sleep trait (GWAS traits uncorrelated with sleep; including GWAS for Alzheimer disease, type 2 diabetes, autism, rheumatoid arthritis and height). H3K9, histone H3 lysine 9.

Full size image

Sleep duration, insomnia symptoms, excessive daytime sleepiness and chronotype are significantly correlated, both at the phenotype and genetic levels (Fig. 1), with greater pairwise correlations in males than in females (Supplementary Fig. 1). Thus, to find loci common to sleep traits, we performed a multiple-trait GWAS51. We identified two new association signals near HCRTR2 and PATJ (INADL) and found that PAX8 and MEIS1 associations influenced multiple sleep traits (Fig. 2, Table 2 and Supplementary Fig. 7). HCRTR2 encodes hypocretin receptor 2, the primary receptor of two receptors for wake-promoting orexin neuropeptides52 involved in narcolepsy and regulation of sleep. Notably, the minor allele at rs3122163 (C) showed subthreshold association with shorter sleep duration and morningness chronotype, suggesting gain of function, but no association with insomnia symptoms. Assessment of objective sleep measures and functional and physiological follow-up should yield important insights into orexin receptor signaling, a pathway important for the pharmacological treatment of narcolepsy53 and insomnia54. PATJ encodes a membrane protein involved in the formation of tight junctions and is implicated in photoreception in mice and Drosophila55,56. The INADL protein is reported to interact with HTR2A (ref. 57), a serotonin receptor with a known role in sleep regulation58,59.

Table 2 Genome-wide significant (P < 5 × 10-8) loci associated with a multiple-phenotype model of sleep duration, insomnia symptoms, excessive daytime sleepiness and categorical chronotype in subjects of European ancestry in UK Biobank
Full size table

Our strongest association for insomnia symptoms fell within MEIS1, a locus previously associated with RLS in GWAS60. Our lead SNP, rs113851554, and the correlated 3′ UTR variant rs11693221 (pairwise r2 = 0.69, D′ = 0.90 in 1000 Genomes Project European (EUR) individuals) represent the strongest known genetic risk factor for RLS and were identified in MEIS1 sequencing studies61,62 following up the original RLS GWAS signal (rs2300478)60,63. Conditional analysis suggests that only one underlying signal, detected by the lead SNP rs113851554 in our GWAS, explains the association of all three SNPs with insomnia symptoms (Supplementary Fig. 8 and Supplementary Table 10). To further investigate the extent of overlap between RLS and insomnia symptoms, we tested a weighted genetic risk score (GRS) for RLS64,65 and found that it was also associated with insomnia symptoms with concordant direction of allelic effects (OR (95% CI) = 1.06 (1.05–1.07) per RLS risk allele, P = 1.17 × 10−21; Supplementary Table 11). Weighting of the RLS GWAS alleles by SNP effect on periodic limb movements (PLMs) did not substantially alter the overall results (Supplementary Table 11). Interestingly, recent data indicating that thalamic glutamatergic activity is increased in RLS provide evidence of an underlying propensity for hyperarousal in RLS66, which is also a core feature of insomnia. Future analyses of pairwise bidirectional causal effects for all three traits will be necessary to determine whether shared genetic associations correspond to causality, partial mediation or pleiotropy.

Strong epidemiological associations of sleep duration, insomnia symptoms and daytime sleepiness have been observed with disease traits, but the extent to which the underlying genetics are shared is unknown. Therefore, we tested for genome-wide genetic correlation between findings from our sleep GWAS and those from publicly available GWAS for 20 phenotypes spanning a range of cognitive, neuropsychiatric, anthropometric, cardiometabolic and autoimmune traits using LD score regression67 (Fig. 4 and Supplementary Table 12).

Figure 4: Genetic architectures shared between sleep duration, insomnia symptoms or excessive daytime sleepiness and 20 behavioral and disease traits.
figure 4

LD score regression67 estimates of genetic correlation (rg) were obtained by comparing our GWAS findings for sleep duration, insomnia symptoms and excessive daytime sleepiness with summary statistics from 20 publicly available GWAS for psychiatric and metabolic disorders, immune diseases and other traits of natural variation. Blue, positive genetic correlation; red, negative genetic correlation; rg values are displayed for significant correlations. Larger squares correspond to more significant P values. Genetic correlations that were significantly different from zero after Bonferroni correction are marked (*P < 1 × 10−3, **P < 1 × 10−5, ***P < 1 × 10−7); the P-value cutoff after Bonferroni correction was 0.0025. All genetic correlations in this report can be found in tabular form in Supplementary Table 12. BMI, body mass index; BMD, bone mineral density, HOMA-IR, homeostatic model assessment of insulin resistance.

Full size image

Genetic correlations demonstrated a strong biological link between longer sleep duration and risk of schizophrenia (rg = 0.29, P = 1 × 10−13), as suggested by previous reports18,47,68. Furthermore, a schizophrenia GRS (96 variants) was associated with longer sleep duration (β (s.e.) = 1.44 (0.36) min/allele, P = 2.56 × 10−4) (interquartile range = 2.3 h), although a variety of sleep behaviors are seen in patients with schizophrenia69,70,71. Significant genetic correlation between low birth weight and longer sleep duration (rg = −0.27, P = 1 × 10−4) may reflect shared links between genetically determined insulin secretion or action pathways underlying fetal growth72,73 and long sleep duration. In support of this hypothesis, significant genetic correlation was observed by Jones et al.18 between oversleeping and both fasting insulin and risk of type 2 diabetes in UK Biobank. Genetic correlation between sleep duration and Crohn's disease risk (rg = 0.18, P = 1 × 10−3) is also consistent with epidemiological observations74.

Significant genetic correlation was also found between increased insomnia symptoms and major depression, adverse glycemic traits, increased adiposity and fewer years of education and between excessive daytime sleepiness and increased adiposity (all P < 1 × 10−3), further highlighting biological overlap of sleep traits with metabolic traits, psychiatric traits and educational attainment17. In support of this overlap, studies have shown that experimentally suppressing slow-wave sleep leads to decreased insulin sensitivity and impaired glucose tolerance75,76. Notably, a GRS for fasting insulin was not significantly associated with insomnia symptoms (7 SNPs, OR (95% CI) = 1.01 (0.99–1.03), P = 0.51). Finally, consistent with a well-established but poorly understood link between excessive daytime sleepiness and obesity77,78, a GRS for BMI was associated with excessive daytime sleepiness (95 SNPs, β (s.e.) = 0.002 (0.0004) sleepiness category/allele, P = 1.67 × 10−4) but not with insomnia symptoms (OR (95% CI) = 1.00 (0.998–1.003), P = 0.73).

Moving forward, replication and systematic testing of genetic correlations in larger samples will be needed. Notably, genetic correlation testing between insomnia and RLS should be examined but was not possible here because RLS consortium GWAS results were not available. Additionally, identifying causal relationships between genetically correlated traits may be difficult, and findings using Mendelian randomization approaches will need cautious interpretation given potential selection biases in UK Biobank79,80,81.

In summary, in a GWAS of sleep traits, we identified new genetic loci that point to previously unstudied variants that might modulate the hypocretin–orexin system and retinal development and influence genes expressed in the cerebral cortex. Furthermore, genome-wide analysis suggests that sleep traits have underlying genetic pathways in common with neuropsychiatric and metabolic diseases. This work should advance understanding of the molecular processes underlying sleep disturbances and open new avenues of treatment for sleep disorders and related disorders.

Methods

Population and study design.

Study participants were from the UK Biobank study, described in detail elsewhere80,81,82. In brief, the UK Biobank is a prospective study of >500,000 people living in the UK. All people in the National Health Service registry who were aged 40–69 years and living <25 miles from a study center were invited to participate from 2006–2010. In total, 503,325 participants were recruited from over 9.2 million mailed invitations. Self-reported baseline data were collected by questionnaire, and anthropometric assessments were performed. For the current analysis, individuals of non-white ethnicity were excluded to avoid confounding effects. All participants provided informed consent to the UK Biobank.

Sleep quality, sleep quantity and covariate measures.

Study subjects self-reported sleep duration, insomnia symptoms, excessive daytime sleepiness, depression, medication use, age, sex, height and weight on a touchscreen questionnaire. For sleep duration, subjects were asked, “About how many hours sleep do you get in every 24 hours? (please include naps),” with responses in hour increments. To assess insomnia symptoms, subjects were asked, “Do you have trouble falling asleep at night or do you wake up in the middle of the night?” with responses “never/rarely,” “sometimes,” “usually” and “prefer not to answer.” To assess daytime sleepiness, subjects were asked, “How likely are you to doze off or fall asleep during the daytime when you don't mean to? (e.g. when working, reading or driving),” with responses “never/rarely,” “sometimes,” “often,” “all the time,” “don't know” and “prefer not to answer.” Approximately 500,000 subjects answered these questions, but only the 120,286 unrelated individuals with genetic data and European ancestry were considered for analysis. Subjects with self-reported shift work (n = 6,557) or sleep medication use (n = 1,184) were excluded. Subjects who responded “do not know” or “prefer not to answer” were set to missing. Sleep duration and excessive daytime sleepiness were untransformed and treated as continuous variables, with daytime sleepiness coded as 1–4. The insomnia symptoms trait was dichotomized into controls (“never/rarely”) and cases (“usually”). Covariates used in sensitivity analyses included self-reported sleep apnea, BMI, depression, psychiatric medication use, socioeconomic, smoking, employment and marital status, and snoring, and secondary GWAS for sleepiness included adjustment for BMI or depression. Sleep apnea cases were defined on the basis of ICD-10 diagnosis code (391 cases). BMI at baseline visit was calculated from entries of height and weight (n = 75,540 with available data). Depression was reported in answer to the question “How often did you feel down, depressed or hopeless mood in last 2 weeks?” (cases (n = 4,242) answered “more than half the days” or “nearly every day”). Medication use was self-reported as part of the initial UK Biobank interview. Our list of psychiatric medications for sensitivity analysis included the four most widely used: fluoxetine (Prozac), citalopram (Cipranol), paroxetine (Seroxat) and sertraline (Lustral). Our list of sleep medications included the 21 most widely used sleep medications in the UK Biobank (oxazepam, meprobamate, medazepam, bromazepam, lorazepam, clobazam, chlormezanone, temazepam, nitrazepam, lormetazepam, diazepam, zopiclone, triclofos, methyprylone, prazepam, triazolam, ketazolam, dichloralphenazone, clomethiazole, zaleplon and butobarbital). Smoking status was self-reported as past and current smoking behavior, and individuals were classified into those with “current” or “past” smoking or “never” smokers. Socioeconomic status was represented by the Townsend deprivation index, on the basis of national census data obtained immediately preceding participation in UK Biobank. Employment status was self-reported (cases were retired; controls were currently employed). Marital status was derived from self-reported household occupancy and relatedness data. Snoring was reported in answer to the question “Does your partner or a close relative or friend complain about your snoring?”

Genotyping, quality control and imputation.

Of the ∼500,000 subjects with phenotype data in the UK Biobank, ∼153,000 are currently genotyped. Genotyping was performed by the UK Biobank, and genotyping, quality control and imputation procedures are described in detail at the UK Biobank website (http://biobank.ctsu.ox.ac.uk/). In brief, blood, saliva and urine were collected from participants, and DNA was extracted from buffy coat samples. Participant DNA was genotyped on two arrays, UK BiLEVE and UKB Axiom, with >95% common content. Genotypes were called using Affymetrix Power Tools software. Sample and SNP quality control were performed. Samples were removed for high rates of missingness or heterozygosity (480 samples), short runs of homozygosity (8 samples), relatedness (1,856 samples) and sex mismatches (191 samples). Genotypes for 152,736 samples passed sample quality control (∼99.9% of total samples). SNPs were excluded if they did not pass quality control filters across all 33 genotyping batches. Batch effects were identified through frequency and Hardy–Weinberg equilibrium tests (P < 1 × 10−12). Before imputation, 806,466 SNPs passed quality control in at least one batch (>99% of the array content). Population structure was captured by principal-component analysis on the samples using a subset of high-quality (missingness <1.5%), high-frequency (>2.5%) SNPs (∼100,000 SNPs) and identified the subsample of European descent. Imputation of autosomal SNPs was performed to a merged reference panel comprising the Phase 3 1000 Genomes Project and UK10K panels using IMPUTE2 (ref. 83). Data were prephased using SHAPEIT3 (ref. 84). In total, 73,355,677 SNPs, short indels and large structural variants were imputed. X-chromosome data were imputed separately, using Eagle 2.0 for prephasing with the --X chromosome flag (no reference panel) for the entire cohort85 and IMPUTE2 with the Phase 3 1000 Genomes Project reference panel for imputation with the --chrX flag on 500-kb chunks in randomly assigned subsets of 30,000 individuals. Post-imputation quality control was performed as previously outlined (http://biobank.ctsu.ox.ac.uk/), and an imputation INFO score cutoff of 0.8 was applied. For GWAS, we further excluded SNPs with minor allele frequency (MAF) <0.001, per-SNP missingness >10% and per-sample missingness >40%. In total, up to 112,586 samples of European descent with high-quality genotyping and complete phenotype/covariate data were used for these analyses.

Statistical analysis.

Phenotypic correlation analysis was performed with the Spearman test in R using the Hmisc package. Genetic association analysis for autosomes was performed in SNPTEST86,87 with the 'expected' method using an additive genetic model adjusted for age, sex, ten principal components and genotyping array. Genome-wide association analysis was performed separately for sleep duration, insomnia symptoms and excessive daytime sleepiness with a genome-wide significance threshold of 5 × 10−8 for each GWAS. We had 80% power to detect the following effects: sleep duration, β = 0.045 h (2.7 min); insomnia symptoms, OR = 1.07; excessive daytime sleepiness, β = 0.021 units (assuming MAF = 0.1, P = 5 × 10−7); we had 80% power to detect the following effects: sleep duration, β = 0.048 h (2.9 min); insomnia symptoms, OR = 1.08; excessive daytime sleepiness, β = 0.023 units (assuming MAF = 0.1, P = 5 × 10−8). X-chromosome analysis was performed in PLINK 1.9 (ref. 88) using linear/logistic regression with separate analysis of the pseudoautosomal regions using the split-chromosome flag, adjusting for sex, age, ten principal components and genotyping array. For the X-chromosome signal at rs73536079, we verified using principal-component analysis that all carriers of the minor allele fell within the major European-ancestry cluster. Follow-up analyses of genome-wide suggestive and significant loci in the primary analyses included covariate sensitivity analyses individually adjusting for sleep apnea, depression, psychiatric medication use, socioeconomic, smoking, employment and marital status, and snoring or BMI (on top of the baseline model adjusting for age, sex, ten principal components and genotyping array). Sensitivity analysis was conducted only in the subset of subjects with data for all secondary covariates (n = 75,477 for sleep duration, n = 39,812 for insomnia symptoms and n = 75,640 for excessive daytime sleepiness). Analysis of enrichment for disease-associated gene sets and transcription factors was performed in WebGestalt46 using the human genome as the reference set, Benjamini–Hochberg adjustment for multiple testing and a minimum of two genes per category. Sex-specific GWAS were performed in PLINK 1.9 using linear/logistic regression stratified by sex adjusting for age, ten principal components of ancestry and genotyping array. We used a hard-call genotype threshold of 0.1 (calls with greater than 0.1 were treated as missing), a SNP imputation quality threshold of 0.80 and a MAF threshold of 0.001. Regional association plots were generated with LocusZoom using the hg19 Nov2014 EUR reference panel to determine background linkage disequilibrium89.

Trait heritability was calculated as the proportion of trait variance due to additive genetic factors across the autosomes, measured in this study using BOLT-REML48 to leverage the power of raw genotype data together with data for low-frequency variants (MAF ≥ 0.001). For multiple-trait genome-wide association analysis, we applied the CPASSOC package developed by Zhu et al.51 to combine association evidence for chronotype, sleep duration, insomnia symptoms and excessive daytime sleepiness. CPASSOC provides two statistics, SHom and SHet. SHom is similar to statistics generated by the fixed-effects meta-analysis method90 but accounts for the correlation of summary statistics due to trait correlation. SHom uses the sample size for a trait as a weight instead of variance, so that it is possible to combine traits with different measurement scales. SHet is an extension of SHom, but power can be improved when the genetic effect sizes are different for different traits. The distribution of SHet values under the null hypothesis of no association was obtained through an estimated beta distribution. To calculate the SHom and Shet statistics, a correlation matrix is required to account for the correlation among traits or resulting from overlapping or related samples from different cohorts. In this study, we directly provide the correlation matrix calculated from the residuals of four sleep traits after adjusting for age, sex, principal components of ancestry and genotyping array. Post-GWAS genome-wide genetic correlation analysis by LD score regression (LDSC)67 was conducted using all UK Biobank SNPs also found in HapMap 3 (ref. 89) and included publicly available data from 20 published GWAS, using a significance threshold of P = 0.0026 after Bonferroni correction for all 20 tests performed. As expected, the observed heritability estimates from LDSC67 using summary statistics for HapMap 3 were lower (5.7% (0.0065%) for sleep duration, 13.3% (0.0123%) for insomnia symptoms and 5.3% (0.005%) for sleepiness) than those calculated by BOLT-REML48 using primary data (10.3% (0.006%) for sleep duration, 20.6% (0.011%) for insomnia symptoms and 8.4% (0.006%) for sleepiness) because the HapMap 3 panel was restricted to variants with MAF >5%. LDSC estimates genetic correlation between two traits (ranging from −1 to 1) from summary statistics using the facts that the GWAS effect size estimate for each SNP incorporates the effects of all SNPs in linkage disequilibrium with that SNP, SNPs with high linkage disequilibrium have higher χ2 statistics than SNPs with low linkage disequilibrium, and a similar relationship is observed when single-study test statistics are replaced with the product of the z scores from two studies of traits with some correlation67. Furthermore, genetic correlation is possible between case–control studies and studies of quantitative traits, as well as within these study types. We performed a weighted GRS analysis using risk scores for RLS, schizophrenia, BMI and fasting insulin. Risk score SNPs passed the genome-wide significance threshold (P < 5 × 10−8) in recent large-scale GWAS and were present in the UK Biobank (RLS, 7 SNPs65, Supplementary Table 11; schizophrenia, 96 SNPs91; BMI, 95 SNPs92; fasting insulin, 7 SNPs93). Independent SNPs were identified, and β estimates were recorded for calculation of the weighted risk score. The GRS was calculated by summing the products of the risk allele count multiplied by the effect reported in the discovery GWAS. The additive genotype model was used for all SNPs. We performed partitioning of heritability using the 25 precomputed functional annotations available through LDSC, which were curated from large-scale robust data sets50. Enrichment both in functional regions and in an expanded region (+500 bp) around loci corresponding to each functional class was calculated to prevent the estimates from being biased upward by enrichment in nearby regions. The multiple-testing threshold was determined using conservative Bonferroni correction (P = 0.05/25 classes).

Data availability.

Summary GWAS statistics will be made available at the UK Biobank website (http://biobank.ctsu.ox.ac.uk/).