GWAS in 446,118 European adults identifies 78 genetic loci for self-reported habitual sleep duration supported by accelerometer-derived estimates

Sleep is an essential homeostatically-regulated state of decreased activity and alertness conserved across animal species, and both short and long sleep duration associate with chronic disease and all-cause mortality1,2. Defining genetic contributions to sleep duration could point to regulatory mechanisms and clarify causal disease relationships. Through genome-wide association analyses in 446,118 participants of European ancestry from the UK Biobank, we discover 78 loci for self-reported sleep duration that further impact accelerometer-derived measures of sleep duration, daytime inactivity duration, sleep efficiency and number of sleep bouts in a subgroup (n=85,499) with up to 7-day accelerometry. Associations are enriched for genes expressed in several brain regions, and for pathways including striatum and subpallium development, mechanosensory response, dopamine binding, synaptic neurotransmission, catecholamine production, synaptic plasticity, and unsaturated fatty acid metabolism. Genetic correlation analysis indicates shared biological links between sleep duration and psychiatric, cognitive, anthropometric and metabolic traits and Mendelian randomization highlights a causal link of longer sleep with schizophrenia.


Population and study design 316
Study participants were from the UK Biobank study, described in detail elsewhere 51 . In 317 brief, the UK Biobank is a prospective study of >500,000 people living in the United 318 Kingdom. All people in the National Health Service registry who were aged 40-69 and 319 living <25 miles from a study center were invited to participate between 2006-2010. In 320 total 503,325 participants were recruited from over 9.2 million invitations. Extensive 321 phenotypic data were self-reported upon baseline assessment by participants using 322 touchscreen tests and questionnaires and at nurse-led interviews. Anthropometric 323 assessments were also conducted and health records were obtained from secondary 324 care data from linked Hospital episode statistics (HES) obtained up until 04/2017. For the current analysis, 24,533 individuals of non-white ethnicity (as defined in "Genotyping 326 and quality control") were excluded to avoid confounding effects. 327 328

Sleep duration and covariate measures 329
Study participants (n ~500,000) self-reported sleep duration at baseline assessment. 330 Participants were asked, "About how many hours sleep do you get in every 24 hours? 331 (please include naps)," with responses in hour increments. Sleep duration was treated 332 as a continuous variable and also categorized as either short (6 hours or less), normal 333 (7 or 8 hours), or long (9 hours or more) sleep duration. Extreme responses of less than 334 3 hours or more than 18 hours were excluded 23 and "Do not know" or "Prefer not to 335 answer" responses were set to missing. Participants who self-reported any sleep 336 medication (see Supplementary Method 1) were excluded. Furthermore, participants 337 who self-reported any shift work or night shift work or those with prevalent chronic 338 disease (i.e., breast, prostate, bowel or lung cancer, heart disease or stroke) or 339 psychiatric disorders (see Supplementary Method 2) were later additionally excluded 340 in a secondary GWAS. 341 Participants further self-reported age, sex, caffeine intake (self-reported cups of 342 tea per day and cups of coffee per day), daytime napping ("Do you have a nap during 343 the day?"), smoking status, alcohol intake frequency (never, once/week, 2-3 344 times/week, 4-6 times/week, daily), menopause status, and employment status during 345 assessment. Socio-economic status was represented by the Townsend deprivation 346 index based on national census data immediately preceding participation in the UK 347 Biobank. Weight and height were measured and body-mass index (BMI) was calculated 348 as weight (kg) / height 2 (m 2 ). Cases of sleep apnea were determined from self-report during nurse-led interviews or health records using International Classification of 350 Diseases (ICD)-10 codes for sleep apnea (G47.3). Cases of insomnia were determined 351 from self-report to the question, "Do you have trouble falling asleep at night or do you 352 wake up in the middle of the night?" with responses "never/rarely", "sometimes", 353 "usually", "prefer not to answer". Participants who responded "usually" were set as 354 insomnia cases, and remaining participants were set as controls. Missing covariates 355 were imputed by using sex-specific median values for continuous variables (i.e., BMI, 356 caffeine intake, alcohol intake, and Townsend index), or using a missing indicator 357 approach for categorical variables (i.e., napping, smoking, menopause, and 358 employment). 359 360

Activity-monitor derived measures of sleep 361
Actigraphy devices (Axivity AX3) were worn 2.8 -9.7 years after study baseline by

Genotyping and quality control 416
Phenotype data is available for 502,631 subjects in the UK Biobank. Genotyping was 417 performed by the UK Biobank, and genotyping, quality control, and imputation 418 procedures are described in detail here 57 . In brief, blood, saliva, and urine was collected 419 from participants, and DNA was extracted from the buffy coat samples. Participant DNA was genotyped on two arrays, UK BiLEVE and UK Biobank Axiom with >95% common 421 content and genotypes for ~800,000 autosomal SNPs were imputed to two reference 422 panels. Genotypes were called using Affymetrix Power Tools software. Sample and 423

Replication and meta-analyses with self-reported sleep duration GWAS 482
Using publicly available databases, we conducted a lookup of lead self-reported sleep 483 duration signals in self-reported sleep duration GWAS results from adult (CHARGE) and 484 childhood/adolescent (EAGLE). If lead signal was unavailable, a proxy SNP was used 485 instead. In addition, we combined self-reported sleep duration GWAS results from adult 486 (CHARGE) and childhood/adolescent (EAGLE) with the UK Biobank (primary model) in 487 fixed-effects meta-analyses using the inverse variance-weighted method in METAL 61 .

Gene, pathway and tissue-enrichment analyses 499
Gene-based analysis was performed using Pascal 39 . Pascal gene-set enrichment 500 analysis uses 1,077 pathways from KEGG, REACTOME, BIOCARTA databases, and a 501 significance threshold was set after Bonferroni correction accounting for 1,077 pathways 502 tested (P <0.05/1,077). Pathway analysis was also conducted using MAGMA 38 gene-503 set analysis in FUMA 62 , which uses the full distribution of SNP P values and is 504 performed for curated gene sets and GO terms obtained from MsigDB (total of 10,891 505 pathways). A significance threshold was set after Bonferroni correction accounting for 506 all pathways tested (P <0.05/10,891). Tissue enrichment analysis was conducted using 507 FUMA 62 for 53 tissue types, and a significance threshold was set following Bonferroni 508 correction accounting for all tested tissues (P <0.05/53). Integrative transcriptome-wide 509 association analyses with GWAS were performed using the FUSION TWAS package 44 510 with weights generated from gene expression in 9 brain regions and 2 tissues from the 511 GTEx consortium (v6). Tissues for TWAS testing were selected from the FUMA tissue 512 enrichment analyses and here we present significant results that survive Bonferroni 513 correction for the number of genes tested per tissue and for all 11 tissues. 514 515

Genetic correlation analyses 516
Post-GWAS genome-wide genetic correlation analysis of LD Score Regression 517 (LDSC) 63-65 using LDHub was conducted using all UK Biobank SNPs also found in 518 HapMap3 and included publicly available data from 224 published genome-wide 519 association studies, with a significance threshold after Bonferroni correction for all tests 520 performed (P <0.05/224 tests). LDSC estimates genetic correlation between two traits 521 from summary statistics (ranging from -1 to 1) using the fact that the GWAS effect-size 522 estimate for each SNP incorporates effects of all SNPs in LD with that SNP, SNPs with 523 high LD have higher X 2 statistics than SNPs with low LD, and a similar relationship is 524 observed when single study test statistics are replaced with the product of z-scores from 525 two studies of traits with some correlation . Furthermore, genetic correlation is possible 526 between case/control studies and quantitative traits, as well as within these trait types. 527 We performed partitioning of heritability using the 8 pre-computed cell-type regions, and 528 25 pre-computed functional annotations available through LDSC, which were curated 529 from large-scale robust datasets 63 . Enrichment both in the functional regions and in an 530 expanded region (+500bp) around each functional class was calculated in order to 531 prevent the estimates from being biased upward by enrichment in nearby regions. The 532 multiple testing threshold for the partitioning of heritability was determined using the 533 conservative Bonferroni correction (P <0.05/25 classes). Summary GWAS statistics will 534 be made available at the UK Biobank web portal. 535 536

Mendelian randomization analyses 537
MR analysis was carried out using MR-Base 538 (https://www.biorxiv.org/content/early/2016/12/16/078972), using the inverse variance weighted approach as our main analysis method 66 , and MR-Egger 67 and weighted 540 median estimation 68 as sensitivity analyses. MR results may be biased by horizontal 541 pleiotropy -i.e. where the genetic variants that are robustly related to the exposure of 542 interest (here sleep duration) independently influence levels of a causal risk factor for 543 the outcome. IVW assumes that there is either no horizontal pleiotropy, or that, across 544 all SNPs, horizontal pleiotropy is (i) uncorrelated with SNP-risk factor associations and 545 (ii) has an average value of zero. MR-Egger assumes (i) but relaxes (ii) by explicitly 546 estimating the non-zero mean pleiotropy, and adjusting the causal estimate accordingly. 547 Estimation of the pleiotropy parameter means that the MR-Egger estimate is generally 548 far less precise than the IVW estimate. The weighted median approach is valid if less 549 than 50% of the weight is pleiotropic (i.e. no single SNP that contributes 50% of the 550 weight or a number of SNPs that together contribute 50% should be invalid because of 551 horizontal pleiotropy). Given these different assumptions, if all three methods are 552 broadly consistent this strengthens our causal inference. For all our MR analyses we 553 used two-sample MR, in which, for all 78 GWAS hits identified in this study for sleep 554 duration, we looked for the per allele difference in odds (binary outcomes) or means 555 (continuous) with outcomes from summary publicly available data in the MR-Base 556 platform. Results are therefore a measure of 'longer sleep duration' and sample 1 is UK 557 Biobank (our GWAS) and sample 2 a number of different GWAS consortia covering the 558 outcomes we explored (Supplementary Table 31,32). The number of SNPs used in 559 each MR analysis varies by outcome from because of some SNPs (or proxies for them) 560 not being located in the outcome GWAS.  .59 x 10 -6 ). For each significant pathway, respective sleep genes are indicated with a colored orange box. Sleep genes from significant pathways that overlap with remaining pathways are indicated in blue. B) Pathway analysis is based on Pascal (gene-set enrichment analysis using 1,077 pathways from KEGG, REACTOME, BIOCARTA databases) Top 10 pathways are shown, and significant pathways are indicated in orange (P <4.64 x 10 -5 ). C) MAGMA tissue expression analysis using gene expression per tissue based on GTEx RNA-seq data for 53 specific tissue types. Significant tissues are shown in red (P <9.43 x 10 -4 ). All pathway and tissue expression analyses in this figure can be found in tabular form in Supplementary Tables 23,24,25.   Figure 3. Genetic architecture shared between sleep duration and behavioral and disease traits. LD score regression estimates of genetic correlation (r g ) were obtained by comparing genome-wide association estimates for sleep duration with summary statistics estimates from 224 publically available GWAS. Blue, positive genetic correlation; red, negative genetic correlation; rg values are displayed for significant correlations. Larger colored squares correspond to more significant P values, and asterisks indicate significant (P <2.2 x 10 -4 ) genetic correlations. All genetic correlations in this report can be found in tabular form in Supplementary Table 27.  Genetic risk scores for sleep duration, short sleep and long sleep were tested using the weighted genetic risk score calculated by summing the products of the sleep trait risk allele count for all 78, 27, or 8 genome-wide significant SNPs multiplied by the scaled effect from the primary GWAS using the GTX package in R. Effect estimates (Beta or OR) are reported per each additional effect allele for sleep duration, short sleep, or long sleep. Abbreviations: CI=confidence interval, GRS=genetic risk score, OR=odds ratio. # Self-reported and varied by cohorts, typically: "How many hours of sleep do you usually get at night (or your main sleep period)?" * In all cohorts, except in GLAKU, child sleep duration was assessed by a single, parent-rated, open question, "How many hours does your child sleep per day including naps?" In GLAKU, parents were asked about the usual bed and rise times during school days, from which the total sleep duration could be estimated.