Main

The UK Biobank (UKB) is a prospective population-based study of more than 500,000 individuals with extensive and readily accessible phenotypic and genetic data1. The release of genome-wide genotyping array data2 for study participants has accelerated genomic discovery through association studies and enabled advances in population genetic analyses, the exploration of genetic overlap between traits and Mendelian randomization studies3,4. Although array data in combination with genotype imputation capture the spectrum of common genetic variants, rare variation that is more likely to modify protein sequences and have large phenotypic consequences is captured less well by these approaches.

Here, we extend the UKB resource with the first study of whole-exome sequencing (WES) for 49,960 UKB participants, generated by the Regeneron Genetics Center, as part of a collaboration with GlaxoSmithKline. These data are available to approved researchers through the UKB Data Showcase (http://ukbiobank.ac.uk/). WES enables the direct assessment of protein-altering variants, the functional consequences of which are more readily interpretable than non-coding variants, providing a clearer path towards mechanistic and therapeutic insights, as well as having a potential use in therapeutic target discovery and validation5,6,7,8 and in precision medicine9,10. Here, we provide an overview of sequence variation in UKB exomes, review predicted damaging variants and their consequences in the general population and perform comprehensive predicted LOF burden testing with 1,730 phenotypes, illustrating the use of this resource in studies of common and rare phenotypes with a focus on deleterious coding variation.

Demographics and clinical characteristics

We selected 50,000 participants from the UKB resource and prioritized individuals with more complete phenotype data: those with whole-body magnetic resonance imaging (MRI) data from the UKB Imaging Study, enhanced baseline measurements, hospital episode statistics (HES) and linked primary healthcare records. Additionally, we selected one disease area for enrichment, including individuals with admission to hospital with a primary diagnosis of asthma (International Clinical Diagnosis 10 (ICD10) codes J45 or J46). This resulted in 8,250 participants with asthma, or approximately 16% of sequenced participants, compared with an incidence of 13% among all 502,543 UKB participants (Table 1). During data generation, samples from 40 participants were excluded due to failed quality-control measures or participant withdrawal, which resulted in a final set of 49,960 individuals of predominantly European ancestry. All study participants provided informed consent. The sequenced participants are representative of the overall 502,543 UKB participants (Table 1) for age, sex and ancestry. Owing to the ascertainment strategy, sequenced participants were more likely to have HES diagnosis codes (84.2% of participants in the WES study compared with 78.0% overall), were enriched for individuals with asthma and had numerous enhanced physical measures (eye measures, hearing tests and electrocardiogram data, Table 1). The exome-sequenced participants did not differ from all participants in the median number of primary and secondary ICD10 codes. The WES individuals include 194 parent–offspring pairs, including 26 mother–father–child trios, 613 full-sibling pairs, 1 monozygotic twin pair and 195 second-degree relatives based on identity-by-descent estimates11 (Supplementary Fig. 1).

Table 1 Clinical characteristics of WES and all UKB participants

Characterization of coding variation in WES data

Exomes were captured using a slightly modified version of the IDT xGen Exome Research Panel v.1.0. The basic design targets 39 megabases of the human genome (19,396 genes among autosomes and sex chromosomes) and was supplemented with additional probes to boost coverage at underperforming loci. In each sample and among targeted bases, coverage exceeds 20× at 94.6% of sites on average (s.d., 2.1%). We observe 4,752,777 variants within targeted regions (Table 2). They include 1,224,107 synonymous (97.8% with minor allele frequency (MAF) < 1%), 2,492,667 missense (98.9% with MAF < 1%) and 198,269 predicted LOF variants that affect at least one coding transcript (initiation codon loss, premature stop codons, splicing and frameshifting insertion or deletion (indel) variants; 99.6% with MAF < 1%); the increasing proportion of rare variants in the LOF and missense categories is consistent with purifying selection (Supplementary Fig. 3a). The median number of variants per individual includes 9,584 synonymous (interquartile range (IQR), 128), 8,702 missense (IQR, 136) and 120 LOF variants (IQR, 12) and is comparable to previous exome-sequencing studies12,13. Restricting analysis to LOF variants that affect all ENSEMBL v.85 transcripts for a gene, the number of LOF variants decreases to 132,336 overall and 95 per individual (a reduction of around 33.2% and about 20.8%, respectively), which is consistent with previous studies. Among LOF variants, we observed similar sequencing depth and allelic balance in comparison to non-LOF variants (Supplementary Table 2). Including non-targeted regions adjacent to exons, we observe 10,189,098 indel and single-nucleotide variants (SNVs) after quality control, 98.5% with a MAF < 1%. As expected, the X chromosome exhibited lower proportions of LOF-to-missense variants (2,492,667 missense and 198,269 LOF variants in autosomes, an approximately 12.6:1 ratio (Table 2); 59,796 missense and 3,670 LOF variants on the X chromosome, an around 16.3:1 ratio) (Supplementary Table 3). This phenomenon is the expected result of the increased efficiency of selection against deleterious variants on the X chromosome.

Table 2 Summary statistics for variants in sequenced exomes of 49,960 UKB participants

Enhancement of coding variation in WES data

To evaluate the enhancements of WES data to the UKB genetic variation resource, we compared the number of targeted autosomal coding variants observed in 49,909 individuals for whom WES, array genotypes and the imputed sequence were all available from the UKB2; therefore, the variant counts for this subset of data differ from Table 2. Variants in the WES dataset were included in this analysis if the data passed the quality-control criteria (see Supplementary Methods) and variants in the imputed sequence dataset were included if the data had an info score > 0.3. Among all autosomal variants, we observed increases in the total number of coding (3,913,039 compared with 647,230), synonymous (1,243,633 compared with 249,263), missense (2,491,290 compared with 384,406) and LOF (198,116 compared with 13,561) variants in the WES dataset compared with the imputed sequence data, respectively (Supplementary Table 4). This represents a 14.6-fold increase in the number of autosomal LOF variants identified by WES compared to those in the imputed sequence. In addition, there was a 16.1-fold increase in the ascertainment of indels in the WES dataset (212,447 (Table 2)) compared to the imputed sequence dataset (13,179) (Supplementary Table 5). Among nearly four-million coding variants observed in the WES data, only 13.0% were also observed in the imputed sequence dataset, highlighting the added value of WES for the ascertainment of rare coding variation.

Concordance of WES, array and imputed genotypes

We summarized the concordance between the WES and array genotypes as well as the WES and imputed allele dosages for the 46,911 individuals of European ancestry for whom all three genetic resources were available, using the squared correlation (R2) of allele counts (see Supplementary Methods). This measure facilitates the interpretation of assessments of accuracy for both rare and common variation14,15. As expected, concordance measured by R2 between genomic measures declines with a decrease in MAF (Supplementary Fig. 4).

R2 concordance between the WES and imputed sequence data ranged from 32.2% for MAF < 0.01% to 95.2% for MAF > 1%, with an average of 53.1% across all allele frequencies. Furthermore, the WES dataset has much greater concordance with array genotypes, as both methods directly assay the variation rather than make a computational prediction. This is particularly true in the rarest allele frequencies for which the accuracy of imputation is limited when using current imputation reference panels. R2 concordance between WES and array genotyped variants was substantially higher, ranging from 73.2% for MAF < 0.01% to 98.7% for MAF > 1%, with an average of 92.3% across all allele frequencies.

LOF variants in WES and imputed sequence data

LOF variants constitute a major class of genetic variation that is of great interest because of their disruption of gene function, their causal role in many Mendelian disorders and the success of leveraging protective LOF variants to identify new drug targets5,6,16. Rare LOF variation is best captured by direct sequencing approaches, such as WES, and we sought to quantify the improved yield of LOF variation from WES compared to array genotyping and the imputed sequence. We compared the number of LOF variants ascertained through WES and the imputed sequence among 46,911 UKB participants of European ancestry. Notably, not all individuals with WES data have data for the imputed sequence available—105 individuals for whom WES data were available failed the quality-control criteria for the genotyping chip and were not included in the UKB genotyping chip or imputed sequence release. We observed a larger number of LOF variants that affected any transcript in the WES dataset compared with the imputed sequence data (248,730 compared with 14,872, respectively) (Supplementary Table 6). Furthermore, we observed a greater number of genes with at least one carrier of a heterozygous LOF variant (17,718 genes from the WES dataset; 7,500 genes from the imputation dataset with an info score > 0.3) and genes with at least one carrier of a homozygous LOF variant (789 from WES; 612 from imputation) (Supplementary Table 7). At equivalent sample sizes (n = 46,911 individuals of European ancestry for whom both WES and imputed sequence data were available), the WES dataset included a greater number of genes with LOF variants, across all carrier count thresholds. Notably, the WES data from 46,911 individuals yielded more genes (17,718) with heterozygous LOF variants than the imputed sequence dataset (with info score > 0.3) for all 462,427 UKB participants of European ancestry (8,724 genes). Tracking the increase in the number of genes with carriers of heterozygous LOF variants with the increase in the number of sequenced samples suggests that we are approaching saturation for this metric, and have probably observed at least one carrier with a heterozygous LOF variant for most of the genes that tolerate these variants, and for most genes overall (Supplementary Fig. 3c). By contrast, the number of genes for which homozygous LOF variants are observed still appears to increase rapidly as more samples in the UKB are sequenced, which suggests that homozygous instances of LOF variants for many more genes can be identified by sequencing additional individuals (Supplementary Fig. 3d).

Projection of variants in 500,000 UKB participants

By extrapolating the data, we can estimate the number of genes for which multiple carriers of LOF variants will be observed once our experiment is complete and all approximately 500,000 participants are sequenced. Cautiously, we currently predict that more than 17,000, 15,000 and 12,000 genes will have at least 10, 50 and 100 carriers of heterozygous LOF variants in the full dataset, respectively (Fig. 1, Supplementary Methods).

Fig. 1: Predicted number of genes in carriers of heterozygous LOF variants in around 500,000 whole-exome sequences from existing WES data.
figure 1

The number of autosomal genes with at least 1, 5, 10, and so on, carriers of heterozygous LOF variants that passed Goldilocks quality control (see Supplementary Methods), had genotype missingness of <10% and Hardy–Weinberg equilibrium P > 10−15 increases with sample size. UKB participants of European ancestry with WES data (n = 46,911) were downsampled at random to the number of individuals specified on the x axis. The number of genes containing at least the indicated count of carriers of heterozygous LOF variants with MAF < 1% as indicated in the legend are plotted on the y axis. The number of autosomal genes is 18,574 in this gene model. The blue dashed line indicates the predicted number of genes (18,273) with at least 1 carrier of a heterozygous LOF variant in 500,000 exomes. Solid curves connect the observed number of genes; dashed curves connect predicted counts from a β-binomial mixture model (see Supplementary Methods).

Comparison to other large-scale WES resources

Our WES results are consistent with those of a recent large-scale survey17 of genetic variation in 141,456 individuals from the Genome Aggregation Database (gnomAD). When we annotate both exome variant lists with the same annotation pipelines and subset the results to similar numbers of individuals and ancestry, we observe 17,718 genes with LOF variants in any transcript (LOFRGC) in 46,911 individuals of European ancestry in the UKB with WES data compared with 17,946 genes with LOF variants in any transcript in 56,885 individuals of non-Finnish European ancestry in gnomAD exomes. Further restricting the data to high-confidence LOF variants with LOFTEE (LOFLOFTEE) (see Supplementary Methods), we obtain 17,323 genes with LOF variants in the UKB and 17,620 in gnomAD (see Supplementary Methods). Consistent with expectation, we observe around 10% more genes with at least 10 LOF variants, including single-nucleotide polymorphisms and indels, and around 12% more genes with at least 10 individuals carrying LOF variants using the LOFRGC definition compared to LOFLOFTEE, respectively (Supplementary Fig. 5).

Survey of medically actionable pathogenic variants

To date, more than 5,000 genetic disorders have known molecular causes and associated genes. The American College of Medical Genetics (ACMG) has proposed 59 genes (ACMG SF2.0)18,19 (hereafter ACMG59) that are associated with highly penetrant disease phenotypes and for which available treatments and/or prevention guidelines can reduce the morbidity and mortality in genetically susceptible individuals. Large-scale human genomic sequencing efforts coupled with electronic health record data provide opportunities to assess penetrance and prevalence of pathogenic and likely pathogenic variants in known monogenic disease genes, as well as investigate the phenotypic effects of variants of unknown significance. In addition, phenotypically agnostic population sampling provides opportunities to better characterize the phenotypic spectrum of these disorders and estimate the associated disease penetrance in the population. Furthermore, these efforts enable the implementation of precision medicine by identifying individuals who carry medically actionable pathogenic variants and providing medical care and surveillance pre-emptively. The informed consent for the UKB does not allow the return of these results to individuals or their healthcare providers as it does in the Geisinger-Regeneron DiscovEHR project10.

We interrogated variant data for the 49,960 individuals with WES from the UKB to identify a ‘strict’ set of reported pathogenic missense and LOF variants (that is, those with at least two stars in ClinVar and no conflicting interpretations) as well as a set of likely pathogenic LOF variants (that is, those in genes for which truncating variants are known to cause disease) according to current ACMG guidelines19 (see Supplementary Methods). We identified 548 such variants (315 reported pathogenic, 233 likely pathogenic) in 992 carriers from the sample of 49,960 individuals (Supplementary Table 8). Nine individuals carried variants in two genes. Of note, 37 of the likely pathogenic variants would qualify as previously reported pathogenic variants using a broader definition (Supplementary Table 9). Pathogenic or likely pathogenic variants were observed in 47 out of 59 genes listed in ACMG59, with a median number of 4 variants per gene and a median of 7 carriers per gene (these calculations included the 12 genes for which there were no observed carriers). Overall, 2.0% of the sequenced individuals carry a reportable pathogenic or likely pathogenic variant in at least one gene from the 59 gene set of ACMG59. Using the same methodology for data from 91,514 participants from the DiscovEHR study20 that have been sequenced to date, we observed a slightly higher percentage of individuals, 2.76%, who carry a potentially actionable rare pathogenic or likely pathogenic variant in the 59 gene set of ACMG59 (Supplementary Table 10). This difference may reflect differences between a study of individuals who seek clinical care (DiscovEHR) and a population-based study not ascertained in the context of active clinical care (UKB).

Among the 47 reportable genes from the 59 gene set of ACMG59, variants in cancer-associated genes were the most prevalent in UKB with WES; BRCA2 (93 variants, 166 carriers), BRCA1 (39 variants, 59 carriers), PMS2 (21 variants, 59 carriers) and MSH6 (35 variants, 52 carriers); and variants in LDLR, which is associated with familial hypercholesterolaemia (Mendelian inheritance in man (MIM) code 143890), were the second most prevalent (35 variants, 68 carriers). Cardiac dysfunction disorders were also highly represented mainly by variants in KCNQ1, which causes long QT syndrome (MIM192500) (25 variants, 55 carriers), PKP2, which causes arrhythmogenic right ventricular dysplasia (MIM 609040) (21 variants, 54 carriers), and MYBPC3, which causes cardiomyopathy (MIM 115197) (25 variants, 50 carriers) (Supplementary Fig. 6a). Of the 992 carriers of medically actionable pathogenic variants, 985 were heterozygous for pathogenic variants in genes that are responsible for dominant conditions. Additionally, we identified six individuals who were homozygous for pathogenic variants in genes associated with autosomal recessive disorders (familial adenomatous polyposis 2, MIM 608456) and mismatch repair cancer syndrome (MIM 276300), which are caused by to biallelic pathogenic variants in MUTYH and PMS2, respectively) and one individual who was hemizygous for an X-linked condition (Fabry disease (MIM 301500), which is caused by pathogenic variants in GLA). Indeed, the male individual who was hemizygous for the pathogenic variant (c.335G>A; p.Arg112His) in GLA has diagnoses of angina pectoris, atrial fibrillation, chest pain and chronic ischaemic heart disease. Similarly, one individual who was homozygous for a pathogenic missense variant (c.1145G>A; p.Gly382Asp) in MUTYH has a history of benign neoplasm of the colon, diverticular disease of the intestine, colonic polyps and intestinal obstruction. These examples illustrate how the extensive health data available for UKB participants provide a valuable resource to assess variant pathogenicity and disease risk at the population level, and the potential to model outcomes for individuals who have pathogenic variants.

We evaluated the penetrance of pathogenic and likely pathogenic variants in BRCA1 and BRCA2 (Supplementary Fig. 6b and Supplementary Table 11) to compare five BRCA1 or BRCA2 associated cancers21 as well as to explore whether risk was conferred for other cancers. Although BRCA1 and BRCA2 have differences in risk among cancer subtypes, due to low sample sizes, we analysed summed counts in 114 male and 110 female participants with reported pathogenic and likely pathogenic variants in BRCA1 or BRCA2. We found that the prevalence of cancers is increased in carriers of pathogenic or likely pathogenic variants in BRCA1 or BRCA2 for five cancers (breast cancer in women, ovarian cancer, prostate cancer, melanoma and pancreatic cancer derived from self-report or HES data) that have previously been associated with mutations in BRCA1 or BRCA2. The cancer prevalence for any of the five cancers was 21.1% in carriers compared with 6.6% in non-carriers; odds ratio (OR) = 3.77 (95% confidence interval (CI) = 2.73–5.21), P = 1.7 × 10−12 (3,337 cases, 46,599 controls)). There was no significant difference between carriers and non-carriers when all other cancers, excluding these five, were combined (15.8% in carriers compared with 17.7% in non-carriers, OR = 0.87 (95% CI = 0.58–1.30), P = 0.55 (8,300 cases, 38,424 controls)). The most-prevalent cancer types in this group were unspecified malignant neoplasms of skin (C443 and C449), and carcinoma in situ of the cervix (D069). Increased risk was observed in the carriers of BRCA1 or BRCA2 variants for each of the five previously associated cancers analysed individually (ovarian cancer in women, OR = 10.07 (95% CI = 4.36–23.29), P = 4.9 × 10−5 (162 cases, 27,076 controls); breast cancer in women, OR = 4.42 (95% CI = 2.80–6.97), P = 2.9 × 10−8 (1,654 cases, 25,584 controls); prostate cancer in men, OR = 3.48 (95% CI = 1.98–6.11), P = 1.5 × 10−4 (888 cases, 21,818 controls); melanoma, OR = 1.9 (95% CI = 0.78–4.62), P = 0.20 (599 cases, 49,345 controls); and pancreatic cancer, OR = 3.48 (95% CI = 1.78–6.82), P = 1.7 × 10−3 (602 cases, 49,342 controls) (P values were calculated using Fisher’s exact tests). These differences in overall cancer risk also manifest as clearly earlier disease onset and rates of cancer-free survival. Comparing the cumulative proportion of female participants who were free of breast and ovarian cancer, we estimate a hazard ratio of 4.31 in carriers of BRCA1 or BRCA2 variants to non-carriers (95% CI = 2.97–6.26, P = 1.4 × 10−16) (Supplementary Fig. 6c). Comparing the cumulative proportion of male participants who were free of prostate cancer, the hazard ratio was 3.68 in carriers of BRCA1 or BRCA2 variants to non-carriers (95% CI = 2.17–6.24, P < 3.3 × 10−7) (Supplementary Fig. 6d). With the UKB WES resource, cancer risk can be more-deeply explored across broader sets of variants as well as with larger WES datasets and continued accrual of incident cancers. Our results corroborate those of another recent population-based application of WES linked to health records10 to evaluate cancer risk in individuals with pathogenic variants in BRCA1 or BRCA2, demonstrating the value of WES to identify high-penetrance rare alleles that are associated with clinical phenotypes; such efforts can be applied across other genetic disorders, enabling the implementation of precision medicine at the population level. We also note that penetrance estimates from population-based studies such as ours appear to be lower than those from previous studies, which may have relied on the screening of high-risk patients or families (for example, we estimate an around 20% risk of breast cancer in carriers of medically actionable variants in BRCA1 or BRCA2 in our sample (median age, 58 years), compared with previous meta-analytic estimates of 30–40% risk at an age of 60 years)22.

Phenotypic associations with LOF variation

The combination of WES, which enables the comprehensive capture of LOF variants, with rich health information allows for the broad investigation of the phenotypic consequences of human genetic variation. We conducted burden tests for rare (MAF < 1%) LOF variants in 16,219 autosomal genes with at least 4 carriers of LOF variants and in 1,730 traits (1,059 discrete traits with at least 50 cases defined by HES and self-reported data; 671 quantitative, anthropometric and blood traits) in 46,876 individuals of European ancestry. For each gene–trait association, we also evaluated the signal for the single variants included in the burden test. We identified 26 unique gene burden–trait associations with P < 10−7; among these, 24 were more significant than any single LOF variant included in the burden test. Given the number of genes and traits analysed, we expected around 2.5 false-positive results to reach P < 10−7 (corresponding to a false-discovery rate of approximately 10%). Reassuringly, the results include several well-established associations (Supplementary Tables 12, 13), in addition to new signals (Table 3), for which we sought replication (Supplementary Tables 14, 15).

Table 3 Novel LOF gene burden associations

For example, we observe that carriers of MLH1 LOF variants, which are associated with Muir–Torre and Lynch syndromes23 (MIM 158320 and 609310, respectively), were at increased risk of malignant neoplasms of the digestive organs (OR = 70, P = 8.2 × 10−11). Carriers of PKD1 LOF variants, the major cause of autosomal dominant polycystic kidney disease24 (MIM 173900), were at increased risk of chronic kidney disease (OR = 86, P = 6.4 × 10−10). Carriers of TTN LOF variants are at increased risk of cardiomyopathy (OR = 12, P = 6.7 × 10−9), which is consistent with previous reports25. In addition to Mendelian disorders, other findings with strong support in the literature include HBB with red-blood-cell phenotypes26, IL33 with eosinophils (driven by rs146597587)27, KALRN with platelet volume (driven by rs56407180)28, GP1BA with platelet volume29, TUBB1 with multiple platelet phenotypes30 and COL4A4 with urine microalbumin31. In some cases, we see patterns of association with traits that may be secondary to known phenotypic associations. For example, ASXL1 and CHEK2 are genes involved in myeloproliferative disorders32 and cancer33, respectively, which may explain the observed associations with haematological traits (which may be secondary to myelodysplastic disease or chemotherapy). Indeed, carriers of LOF variants in ASXL1, a known driver of clonal haematopoiesis of indeterminant potential, are 5 years older on average than non-carriers. This is consistent with the possibility that many of the observed ASXL1 LOF variants are somatic mutations. Of the 103 carriers of ASXL1 LOF variants, 45 (44%) were identified as carriers of LOF variants for clonal haematopoiesis of indeterminant potential (see Supplementary Methods). We observed no significant differences in age for carriers of LOF variants in the other genes included in Tables 3 or Supplementary Tables 12, 13 (Supplementary Table 16).

Many other known phenotypic associations are supported by the data at more-modest significance thresholds (Supplementary Table 13). These include, for example, associations between LOF variants in LDLR with coronary artery disease34, GP1BB with platelet count35, KRT5 with fibrosis and scarring of skin36, as well as PALB2 and BRCA2 with cancer risk37,38. We observed no exome-wide significant LOF burden associations on chromosome X.

LOF associations and novel gene discovery

Our LOF gene burden association analysis identified six gene–trait LOF associations with P < 10−7 that have not previously been reported (Table 3). We identified a novel association between PIEZO1 LOF variants (cumulative allele frequency = 0.18%) and increased risk for varicose veins (OR = 4.9, P = 3.2 × 10−8). This finding is driven by a burden of rare LOF variants, with the most-significant PIEZO1 single-variant LOF association achieving a P value of 2.6 × 10−3. ‘Leave-one-out’ sensitivity analyses indicated that no single variant accounted for the entire signal and step-wise regression analyses indicated that three separate variants (one of which had a minor allele count of greater than one) were contributing to the burden signal (Supplementary Table 17). We replicated this finding for varicose veins (1,572 cases, 75,704 controls) in WES data from the DiscovEHR study (OR = 3.8, P = 1.5 × 10−6) (Supplementary Table 14). This region had previously been implicated in varicose veins by common non-coding variants with small effects39 (rs2911463, OR = 0.996). PIEZO1 encodes a 36-transmembrane-domain cation channel that is highly expressed in the endothelium and has a critical role in the development and adult physiology of the vascular system as it translates shear stress into electrical signals40. This previous report of the rs2911463 variant mapped the association to PIEZO1 through evidence of gene function and analysis using DEPICT, but did not find strong evidence in expression quantitative trait loci that would clarify the mechanism in order to modulate the target therapeutically39. Rare missense variants have been reported in families segregating autosomal-dominant dehydrated hereditary stomatocytosis (MIM 194380), which is characterized by haemolytic anaemia with primary erythrocyte dehydration due to decreased osmotic fragility. Additionally, biallelic LOF variants in PIEZO1 have been reported in families with lymphatic malformation syndrome (MIM 616843), a rare autosomal recessive disorder characterized by generalized lymphoedema, intestinal and/or pulmonary lymphangiectasia and pleural or pericardial effusions. Notably, some of the reported patients with lymphoedema presented with varicose veins and deep-vein thrombosis41. Our data not only provide support for PIEZO1 as the causal gene in this locus, but also clarify the direction of effect, as the loss of gene function in carriers with heterozygous LOF variants leads to an increased risk of developing varicose veins.

We also identified a novel LOF burden association with MEPE (cumulative MAF = 0.18%) and decreased bone mineral density (BMD) (as measured by heel BMD t-score, −0.46 s.d., P = 1.2 × 10−8). Leave-one-out analyses suggested that the aggregate signal is driven by multiple variants (Supplementary Table 18), only one of which could be imputed (rs753138805, which encodes a frameshift that results in an early truncation of the protein). In analysis of all 500,000 UKB participants, rs753138805 was significantly associated with decreased BMD (−0.4 s.d., P = 2.0 × 10−18) and showed a trend for increased risk of osteoporosis (OR = 1.9, P = 0.10, n = 3,495 cases, Supplementary Table 15). These findings are corroborated by analysis of the HUNT study42, in which rs753138805 was associated with decreased BMD (−0.5 s.d., P = 2.1 × 10−18) and increased risk of any fracture (OR = 1.4, P = 1.6 × 10−5, n = 24,155 cases) (Supplementary Table 15). In the DiscovEHR study, we observed a directionally consistent, but nonsignificant, association for MEPE LOF variants with femoral-neck BMD t-score (−0.19, P = 0.22) (Supplementary Table 14). The MEPE locus was previously reported43 to have six independent signals with modest effects on BMD.

Although two of the previously reported non-coding variants are in moderate (r2 = 0.5) or high (r2 = 0.78) linkage disequilibrium with two variants that contribute to the burden test, the burden association is only partially attenuated in conditional analyses (P = 2.9 × 10−4, including all six previously reported variants). MEPE encodes a secreted calcium-binding phosphoprotein that has a key role in osteocyte differentiation and bone homeostasis44,45. Studies in Mepe−/− knockout mice (Mepetm1Tbrw) have yielded inconsistent results, with two groups reporting an increase in BMD45,46 and another reporting no change47.

Another new signal comprises LOF variants in COL6A1 (cumulative allele frequency = 0.03%) that are associated with a 2.7 mm Hg decrease in corneal resistance factor (−1.5 s.d., P = 4.7 × 10−10) and corneal hysteresis (−1.3 s.d., P = 2.9 × 10−8), which are measures of corneal biomechanics48. COL6A1 encodes a component of the collagen type-VI microfibrils, which have important roles in maintaining the structure and function of the extracellular matrix and which are major components of the human cornea49. This locus has previously been implicated in ocular traits; rs73157695 and nearby common variants have been associated with myopia50 (OR = 0.94; P ≈ 10−13) and intraocular pressure51. Leave-one-out analyses indicate that multiple variants are driving the association (Supplementary Table 19) and that these are not in strong linkage disequilibrium with previously reported variants (Supplementary Tables 20, 21). Protein levels of COL6A1 were reduced in eyes from patients with keratoconus52, and individuals with keratoconus and other corneal diseases, such as Fuchs’ corneal dystrophy, have reduced corneal hysteresis and corneal resistance factor53. Measures of corneal hysteresis and corneal resistance factor were not available in the DiscovEHR study for replication analyses.

The new LOF burden associations for IQGAP2, GMPR (driven by rs147049568) and FAM160B1 relate to platelet morphology, haemoglobin levels and wound healing, respectively. Previous genome-wide association and exome-chip studies have implicated IQGAP2 and GMPR in haematological variability28,54. Our results for these two genes, each of which could be replicated in the DiscovEHR study (Supplementary Table 14), provide additional evidence of causal roles for these genes and establishes the direction of effect with respect to gene function on haematological traits. We also observe nominal evidence for replication of FAM160B1 LOF variants with disruption of wound healing in the DiscovEHR study (OR = 13.2, P = 0.02) and find that no single variant drives the association signal in leave-one-out analyses (Supplementary Table 22). However, little is known about the function of FAM160B1, and further characterization is necessary.

In equivalent sample sizes, the LOF burden results from the imputed sequence dataset would not have uncovered the LOF associations at P < 10−7 as identified by WES (Supplementary Table 23). Furthermore, for 19 out of 26 LOF burden results described herein, gene burden results from WES in the 50,000 participant subset were more significant than burden results from the imputed sequence dataset of all 500,000 UKB participants (Supplementary Table 23), demonstrating the value of rare variants captured by WES to power these associations. Notably, we observed qualitatively similar significance for burden tests using LOFRGC and LOFLOFTEE criteria (Supplementary Fig. 7).

Discussion

Integration of large-scale genomic and precision medicine initiatives offer the potential to revolutionize medicine and healthcare. Such initiatives provide a foundation of knowledge that link genomic and molecular data to health-related data at the population scale, enabling researchers to more-completely and systematically study genetic variation and its functional consequences on health and disease. Here, we describe the initial study of large-scale WES data for 49,960 UKB participants, which—to our knowledge—is currently the largest open-access resource of WES data linked to health records and extensive longitudinal study measures. These data greatly extend the current genetic resource, particularly in ascertainment of rare coding variation, which we demonstrate is useful for resolving variant-to-gene links and the directionality of gene-to-phenotype associations.

After quality control, we observed nearly four million SNVs and indel coding variants. Only approximately 10.1% of the coding variants identified by WES were observed in the sequence imputed from the array of 46,911 participants with both WES and imputed sequence, highlighting the added value of WES. This enrichment was even more pronounced with LOF variation for which WES identified more than 240,000 targeted and untargeted LOF variants of which only approximately 4% of these were present in the imputed sequence dataset. Furthermore, 25.9% of the coding variants in the imputed sequence were not observed in the WES data that may represent a large proportion of rare variants that have poor imputation accuracy, as observed in our concordance and visual validation analyses. A small proportion of these variants, seen only in the imputed sequence, also represent variants that are not in the regions targeted by the exome capture design and sequencing, a limitation of the targeted-capture approach. Increasing the numbers of individuals and ancestral diversity in imputation reference panels is expected to improve imputation accuracy for rare variants.

As with previous studies of this size20, we observed a large number of LOF variants, including at least one carrier of a rare heterozygous LOF variant in 97% of autosomal genes (compared to more than 41% of autosomal genes in the imputed sequence data for the same participants). It is important to note that our LOF annotation strategy is geared towards increasing sensitivity for the identification of LOF variants and downstream association discovery. While the number of genes with heterozygous instances of LOF variants is approaching saturation at this sample size, WES in the entirety of the UKB resource will considerably increase the number of carriers of LOF variants and the ability to detect phenotypic associations. We also observe 789 autosomal genes with homozygous instances of rare LOF variants, and this number of genes will also increase with continued sequencing of all UKB participants; however, studies in populations with a high degree of parental relatedness55,56 will provide even more genes with homozygous LOF variants and complement efforts such as the UKB. LOF variation is an extremely important class of variation for identifying drivers of high genetic risk of disease, new disease genes and therapeutic targets. Very large sample sizes are needed to detect new LOF associations given their collective rare allele frequencies. The WES data provide a substantial enhancement to the number of LOF variants identified and the power to detect new associations, which will only improve with continued sequencing of all UKB participants.

We illustrate the unique value of this expanded WES resource in the UKB to assess pathogenic and likely pathogenic variants in a disease-agnostic large-scale population-based study with longitudinal follow-up. We conducted a survey of pathogenic and likely pathogenic variants in medically actionable ACMG59 genes. Using stringent variant-filtering criteria, we arrived at an estimated prevalence of 2% of individuals in this study population having a clinically actionable finding. This resource enables us to characterize disease-risk profiles for individuals who carry pathogenic and likely pathogenic variants in medically relevant disease genes, including cancer susceptibility genes, such as BRCA1 and BRCA2. We observed that carriers of pathogenic or likely pathogenic variants had 3.77-fold greater odds of any of the five cancers previously associated with BRCA1 or BRCA2; penetrance for any of the five cancers was 21.1% in carriers compared with 6.6% in non-carriers. We further explored whether these variants conferred risk to any other cancers and did not observe any such associations. This resource will be valuable for the assessment of variant pathogenicity, particularly for variants of unknown significance and novel variants, and in exploring the full spectrum of disease risk and phenotypic expression. One limitation of the resource for such purposes is limited ancestral diversity. This and other similar studies also highlight the value and potential to apply large-scale sequencing at the population scale to identify a meaningful proportion of individuals who are at high risk of diseases for which effective interventions are available that can reduce the morbidity and mortality of genetically susceptible individuals; such precision medicine approaches could substantially reduce the burden of many diseases.

We conducted gene-burden association testing for LOF variants across all genes and included more than 1,700 binary and quantitative traits. In addition to replication of numerous positive controls, we also highlight previously undescribed significant LOF associations that provide new biological and genetic insights of large effect on disease traits of interest, including PIEZO1 for varicose veins, MEPE for BMD and COL6A1 for corneal thickness. We identified a previously undescribed gene-burden association in PIEZO1, a mechanosensing ion channel that is present in endothelial cells in vascular walls; LOF variants confer a nearly fivefold increased odds of varicose veins in carriers of heterozygous LOF variants in PIEZO1. We also identified a previously undescribed LOF burden association in MEPE, which is associated with decreased BMD and an approximately 2-fold increased odds of osteoporosis and 1.5-fold increased risk of fractures. Overall, through WES and gene-burden tests of association for LOF variants, we identified 26 unique gene–trait associations that exceeded a P < 10−7, of which 24 were more significant than any single LOF variant included in the burden test, highlighting the value of WES and the ability to detect new associations that are driven by rare coding variation. Although several of these regions had previously been identified in genome-wide association studies of more than 10× the sample size, a key strength of the current approach is the identification of likely causal genes and the direction of effect: two key pieces of information required for translation towards new therapeutics. This survey of rare LOF associations was limited by sample size for most binary traits but was well-powered for many quantitative traits. Although surveys of LOF variation in the entire UKB study using array and imputed sequence data have identified LOF associations in previous reports57,58, the WES data identified new associations that are unique to the exome sequence and detected in only approximately one tenth of the sample size; our results highlight the considerable power of WES for the discovery of LOF variants and rare variant associations and the further promise of new biological insights through the sequencing of all participants in the UKB resource.

Efforts are underway to sequence the exomes of all 500,000 UKB participants; these efforts will greatly expand the rare coding variation ascertained, including the number of carriers of heterozygous LOF variants that can now be observed in nearly all genes and the number of genes for which naturally occurring homozygous knockouts can be observed. Coupled with the rich laboratory, biomarker, health-record, imaging and other health-related data that are continually added to the UKB resource, WES will enhance the power for discovery and will continue to yield many important findings and insights. The WES data are available to approved researchers through similar access protocols as existing UKB data (http://ukbiobank.ac.uk/).

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.