A large genome-wide association study of QT interval length utilizing electronic health records

Abstract

QT interval length is an important risk factor for adverse cardiovascular outcomes; however, the genetic architecture of QT interval remains incompletely understood. We conducted a genome-wide association study of 76,995 ancestrally diverse Kaiser Permanente Northern California members enrolled in the Genetic Epidemiology Research on Adult Health and Aging cohort using 448,517 longitudinal QT interval measurements, uncovering 9 novel variants, most replicating in 40,537 individuals in the UK Biobank and Population Architecture using Genomics and Epidemiology studies. A meta-analysis of all 3 cohorts (n = 117,532) uncovered an additional 19 novel variants. Conditional analysis identified 15 additional variants, 3 of which were novel. Little, if any, difference was seen when adjusting for putative QT interval lengthening medications genome-wide. Using multiple measurements in Genetic Epidemiology Research on Adult Health and Aging increased variance explained by 163%, and we show that the ≈6 measurements in Genetic Epidemiology Research on Adult Health and Aging was equivalent to a 2.4× increase in sample size of a design with a single measurement. The array heritability was estimated at ≈17%, approximately half of our estimate of 36% from family correlations. Heritability enrichment was estimated highest and most significant in cardiovascular tissue (enrichment 7.2, 95% CI = 5.7–8.7, P = 2.1e−10), and many of the novel variants included expression quantitative trait loci in heart and other relevant tissues. Comparing our results to other cardiac function traits, it appears that QT interval has a multifactorial genetic etiology.

QT interval length is an important risk factor for adverse cardiovascular outcomes. To enhance the list of QT-interval-associated variants, Hoffmann et al. perform a GWAS of 76,995 ancestrally diverse GERA cohort members followed by meta-analysis with the UKB and PAGE consortia, identifying a total of 28 novel variants. Comparing these results to other cardiac function traits indicates that QT interval has a multifactorial genetic etiology.

Introduction

The QT interval is the time between the onset of ventricular depolarization (Q wave) and completion of repolarization (end of the T wave) in an electrocardiogram (ECG). QT prolongation can have profound prognostic implications for an increased risk of ventricular arrhythmias, most notably Torsades de Pointes (TdP) and sudden cardiac death (Forssell and Orinius 1981; Moller 1981; Taylor et al. 1981; Puddu and Bourassa 1986; Algra et al. 1993; Schwartz et al. 1998), as well as for an increased risk of coronary heart disease and all-cause mortality (Zhang et al. 2011).

The QT interval is heritable (h² ≈ 35–40%) and has multiple environmental and genetic contributors (Dalageorgou et al. 2008; Hodkinson et al. 2016). Previous genome-wide association study (GWAS) (Arking et al. 2006, p. 1; Newton-Cheh et al. 2009; Pfeufer et al. 2009; Nolte et al. 2009; Marroni et al. 2009; Kim et al. 2012; Smith et al. 2012; Jeff et al. 2013; Deng et al. 2013; Arking et al. 2014; Gondalia et al. 2017; Méndez-Giráldez et al. 2017; Bihlmeyer et al. 2018; van Setten et al. 2019; Wojcik et al. 2019; Lahrouchi et al. 2020; van Duijvenboden et al. 2020), the largest with up to 76,061 individuals from 31 cohorts (Arking et al. 2014), have identified ∼36 genome-wide significant (P < 5e−8) loci associated with QT interval length, and build upon prior candidate gene work (Palmer et al. 2012; Seyerle et al. 2014; Kertai et al. 2016). These variants explain only a limited amount of the variance (≈5–10%) (Newton-Cheh et al. 2009; Arking et al. 2014); however, the array heritability has been estimated at 21% (Yang, Manolio, et al. 2011), suggesting many more loci remain. In addition, these studies have not incorporated any adjustments for medications that might putatively impact QT interval length, and thus might be associated with disease targets for these medications, although several pharmacogenetic studies have also been conducted (Volpi et al. 2009; Åberg et al. 2012; Avery et al. 2014; Noordam et al. 2017; Drago and Fischer 2018; Floyd et al. 2018).

We thus reasoned that novel insights into the genetic architecture of QT interval length could be garnered from additional analysis using 3 cohorts. Two of these cohorts had individual level data: Genetic Epidemiology Research on Adult Health and Aging (GERA) and UK Biobank (UKB). The other cohort, Population Architecture using Genomics and Epidemiology (PAGE), had only summary statistics available. First, we tested for QT interval length associations with genetic ancestry principal components (PCs) in the GERA cohort. Second, we utilized the GERA cohort for discovery, with a formal replication test in a meta-analysis of UKB and PAGE study. Third, we maximized locus discovery with a meta-analysis of all 3 (GERA, UKB, and PAGE). Fourth, we conducted sensitivity analysis to determine if medications impacted discovery in GERA. Fifth, we conducted conditional analysis using individual level data (not summary statistics) in GERA, with a replication test using individual level data from the UKB. Finally, we looked at heritability, variance explained by the novel SNPs, and the gain from having multiple measures.

Materials and methods

Figure 1 displays a flowchart describing the study, including the (A) phenotype curation, (B) imputation reference panels, (C) imputation strategy, and (D) analysis and replication workflow. Further details are provided below.

Fig. 1.

Study design. a) Phenotype extraction. b) Imputation reference panel individuals overlap. c) GWAS cohort imputation. SNVs in the HRC were difficult to harmonize. d) GWAS analysis, replication, and meta-analysis strategy. GERA, Genetic Epidemiology Resource on Adult health and aging; UKB, UK Biobank; PAGE, Population Architecture using Genomics and Epidemiology; HRC, haplotype reference consortium; SNV, single nucleotide variant; NHW, non-Hispanic whites; LAT, Latinos; EAS, East Asians; AFR, African Americans; SAN, South Asians; EUR, Europeans.

GERA participants and phenotype

Our primary analysis was of individuals with comprehensive electronic health records from the Kaiser Permanente Research Program on Genes, Environment and Health (RPGEH) GERA cohort, described previously (Banda et al. 2015; Kvale et al. 2015), which are part of Kaiser Permanente Northern California (KPNC). In KPNC, an ECG research database from routine medical care was started in 1995, containing a large number of QT measurements. From this database, we utilized longitudinal QT corrected for heart rate (QT_c; according to Fridericia’s formula) measurements (Fridericia 2003), from 1995 to 2018. The extraction process and detrending adjustment for a secular trend due to an algorithm change (PH7 to PH8) has been described in detail (Mantri et al. 2021). In addition, individuals were excluded if they had a left (LBBB) or right (RBBB) bundle branch block (BBB), identified from the ECGs, or by the following ICD codes: 426.3 (ICD-9 LBBB), 426.4 (ICD-9 RBBB), I44.7 (ICD-10 LBBB), and I45.10 (ICD-10 RBBB).

A total of 66 putative QT-prolonging medications, which were each dispensed to at least 20 members, were identified from crediblemeds.org (accessed 2020 May 20; known, possible, and conditional risk of TdP, and no other filters). Dispensing dates were identified through the KPNC outpatient pharmacy database, which contains most if not all medication use (Karter et al. 2009). QT measurements were considered under the influence of each medication if they were between the initial prescription dispensation date and 7 days after the supply would have ended, except for amiodarone, which was kept until 365 days after the end of days’ supply, due to its long-lasting effect (Hempenius et al. 2019). QT measurements were excluded if they were subject to a rarely used medication (prescribed to <20 users).

KPNC and UCSF institutional review boards approved of this project. Written informed consent was obtained from all subjects.

GERA genotyping, quality control, and imputation

The initial GERA genotyping on 4 Axiom arrays (>650,000 SNPs) optimized for Europeans, Latinos, East Asians, and African Americans, has been described (Hoffmann, Kvale, et al. 2011; Hoffmann, Zhan, et al. 2011). Genotype quality control procedures (Kvale et al. 2015) were redone using an updated normalization step for SNPs that had failed genotyping with the previous algorithm (either marked not recommended in too many batches, or flagged by the outlier plate detector), as has been described (Choquet et al. 2021).

Imputation was performed on an array-wise basis, since each array assayed different variants. Each array was prephased with Eagle v2.3.2 (Loh et al. 2016), and then imputed with Minimac3 v2.0.1 (Das et al. 2016), twice, using 2 different reference panels [each as cosmopolitan reference panels, rather than reference panels that most closely match the genetic ancestry of the target population, as cosmopolitan reference panels have been shown to work well (Howie et al. 2011)], in a manner almost identical to that done in the UKB (Bycroft et al. 2017) (Fig. 1, b and c). The 2-reference panels were (1) the EGA release of the Haplotype Reference Consortium (HRC; n = 27,165) (McCarthy et al. 2016) and (2) the 1000 Genomes Project Phase III release (KGP; n = 2,504; e.g. indels) (Birney and Soranzo 2015). Figure 1b shows the overlap in the panels, the HRC contains all of the KGP individuals, but has less variants in it, particularly indels, as they were left out because they are much harder to harmonize across a large number of studies. Thus we imputed using the HRC, and prioritized imputation from the HRC over KGP, because it generally gives better imputation than the KGP, as has been shown (Bycroft et al. 2017), but also impute from the KGP to obtain additional variants that are not in HRC (e.g. indels). The strategy is virtually identical to that which was done in the UKB (Bycroft et al. 2017), except that the second reference panel in the UKB is a merge of KGP with the UK10K, instead of the KGP used here.

Phenotype genetic ancestry distributions

We smoothed the distribution of covariate-adjusted QT interval length (see GWAS analysis section) over the genetic ancestry PCs (from SNPs with call rate >99.5%) (Banda et al. 2015), using a radial kernel density estimate, as has been described (Hoffmann et al. 2018). We overlaid self-reported race/ethnicity for GERA, and nationality for the HGDP (Banda et al. 2015; Hoffmann et al. 2018).

GERA GWAS analysis and covariate adjustment

We first analyzed each of the 5 GERA groups (non-Hispanic whites, Latinos, East Asians, African Americans, and South Asians, based on self-reported race/ethnicity) separately. For computational efficiency (Hoffmann et al. 2017), and to allow for multiple measurements per individual, and a better estimate for effects of medications with fewer individuals taking them, we first ran a linear mixed model of QT interval length adjusting for sex, age to the third power, and indicator variables for medication (unless otherwise specified) in all GERA groups combined. We then calculated the long-term average residual [by individual (Ganesh et al. 2014)], winsorized those at ±3 SD, and then used those in a leave 1 chromosome out machine-learning regression method of Regenie v2.0.2 (Mbatchou et al. 2021) to account for population substructure and cryptic relatedness, additionally adjusting for genetic ancestry PCs (top 10 for the largest group of non-Hispanic whites and 6 for the other smaller groups) to further adjust for population structure and environmental confounders (Zhang and Pan 2015). We analyzed variants with imputation quality ≥0.3 and minor allele frequency (MAF) ≥0.001, ≥0.002, ≥0.005, ≥0.01, and ≥0.05 for the non-Hispanic whites, Latinos, East Asians, African Americans, and South Asians, respectively, resulting in 15,853,782; 16,948,902; 10,466,294; 16,330,949; and 7,254,556 imputed SNPs included in the analysis (some of these groups have more variants despite a higher MAF, which we suspect may be due to imputation from different arrays with different numbers of typed variants, and the amount of genetic variation in each group). Loci were considered novel if they were >1 Mb from any previously described locus, unless they were within a region of high linkage disequilibrium (LD; e.g. chromosome ends, centromeres, HLA), in which case they were not considered novel. The 5 GERA groups were then combined using a fixed-effects meta-analysis with Metasoft v2.0 (Han and Eskin 2011) and assessed heterogeneity with Cochran’s Q and the I² statistic (Huedo-Medina et al. 2006).

To find additional conditionally independent genome-wide significant SNPs at existing or newly discovered loci, we conducted a forward stepwise regression analysis (again meta-analyzing all 5 GERA groups again), using individual level data, and all SNPs with r_info² > 0.8 within ±1 Mb of the lead SNP, using the same approach as above, but with Regenie v3.1.1.

Replication of newly associated SNPs

To test the novel genome-wide significant results identified in GERA for replication, we evaluated the associations of these variants with QT in a combined UK Biobank (UKB) plus PAGE meta-analysis, using the same approach as in GERA.

UK Biobank

The UKB cohort has been previously described, with imputation as previously described (and as discussed above) (Bycroft et al. 2017). A total of 24,051 individuals with genotype data also had QT measurements. We considered analyzing different self-reported ethnic groups, as in GERA, but only the self-reported whites with global genetic ancestry PC₁ ≤ 70 and PC₂ ≥ −80 [as described in Hoffmann et al. (2017); n = 23,189] had a large enough sample size for stable estimates. QT_c was measured from a 12-lead ECG. Exclusions were similar to before, using ICD10 codes I447 (LBBB) and I451 (RBBB) to identify individuals with BBB. Covariate adjustment was similar as for GERA, here utilizing self-reported medication use. Analyses were otherwise identical to GERA, except with an initial linear model, since there was only 1 QT measure per individual.

We additionally removed rs756835686 as we strongly suspected it was a sequencing error in the UK10K; the variant was imputed only in the UKB, since it had frequency 4.7% in the UKB, but in Europeans in gnomAD (https://gnomad.broadinstitute.org/) the frequency was 0.02% (n = 18,810). The final number of variants tested in UKB with imputation quality ≥0.3 and MAF ≥0.001 was 17,130,898.

Population Architecture using Genomics and Epidemiology

The PAGE study and analysis of the QT phenotype has been described in detail (Wojcik et al. 2019), and consisted of 17,348 non-European individuals (self-reported 44.6% Hispanic/Latino, 34.7% African American 9.4% Asian, 7.9% Native Hawaiian, 1.3% Native American, 2.1% other) on the QT trait, and we have utilized the summary statistics from their paper (see Data availability). The QT interval was calculated from 12-lead ECGs, and models were adjusted for age, sex, study/center, self-identified race/ethnicity, heart rate, and the first 10 genetic ancestry PCs. We note that PAGE did not conduct the same medication adjustment analysis as in GERA and the UKB.

GERA meta-analysis with PAGE and UKB

We performed a meta-analysis of the GERA, UKB, and PAGE together for additional discovery, albeit with no replication cohort, using the same meta-analysis approach described in previous sections.

Sensitivity analysis including vs not including medication

We conducted a sensitivity analysis to assess what effect ignoring medication would have on an analysis using the GERA data. We first conducted a GWAS in GERA using all untreated measurements (52,123 individuals), for the maximum cleanest discovery test. We compared those results with a GWAS using the remainder of the individuals (24,872 individuals) who had only treated measurements (note that the sets of individuals are mutually exclusive).

Replication of loci previously identified

A number of prior studies have identified QT variants, which we tested for replication in the meta-analysis of the 2 studies not previously analyzed (and thus independent of prior discovery), GERA and the UKB. From the previously reported variants (Buniello et al. 2019) from these studies involving QT (Arking et al. 2006, p. 1; Newton-Cheh et al. 2009; Pfeufer et al. 2009; Nolte et al. 2009; Marroni et al. 2009; Kim et al. 2012; Smith et al. 2012; Jeff et al. 2013; Deng et al. 2013; Arking et al. 2014; Gondalia et al. 2017; Méndez-Giráldez et al. 2017; Bihlmeyer et al. 2018; van Setten et al. 2019; Wojcik et al. 2019; Lahrouchi et al. 2020; van Duijvenboden et al. 2020) or pharmacogenetic interactions with QT (Volpi et al. 2009; Åberg et al. 2012; Avery et al. 2014; Noordam et al. 2017; Drago and Fischer 2018; Floyd et al. 2018) we defined an independent locus as we did in GERA above, requiring the lead variant within each locus to be >1 Mb apart from each other (and collapsing long stretches of LD). A total of 110 loci were previously associated with QT interval; however, 36 of these independent loci were previously reported with P ≤ 5e−8 (i.e. genome-wide significant), while 74 were reported with P > 5e−8. We looked at the replication rate of these 2 sets separately, reporting those that meet a Bonferroni correction (P ≤ 0.00045, for 110 tests) and nominal (P ≤.05) significance. We also report on all SNPs previously reported at each locus.

SNP and eQTL annotation

SNPs were annotated using GTEx v8 (GTEx portal on 2022 April 24). Genes were prioritized by first noting that a SNP was an expression quantitative trait locus (eQTL) in heart or artery tissues, then if an eQTL in other tissues, and if not an eQTL, by the closest gene.

Testing for dominance, epistasis, and sex differences in GERA

Within GERA, we conducted 3 additional tests. First, we tested previously and newly reported variants for dominance deviation from additivity; i.e. whether the increase in QT length was linear in the number of alleles an individual had, or if there was a departure from that. The linear regression model was similar to the model used in the GWAS, except that the genetic effect was modeled differently. Namely, we used the best guess genotype (i.e. 0, 1, or 2 copies of the allele, instead of the probabilistic dosage), and included 2 terms in the regression model, an additive term for the genotype, and an additional term for dominance (the coefficient of interest, coded as 1 for both of the homozygote genotypes, and -2 for the heterozygote genotype). We used a Bonferroni correction of P < 0.00032 to correct for the 153 previously and newly reported independent variants.

Second, we tested for epistasis between each pair of the previously and newly reported variants, i.e. if there was a statistical interaction between 2 variants. For each pair, we again fit models similar to that described previously, again using the best guess genotype. We included 3 terms in the model for the genotypes, a main effect of the first variant coded additively (i.e. 0, 1, or 2 copies of the allele), a main effect of the second variant, and an interaction between the 2 of them (i.e. the product of the additive codings of the 2 loci; the coefficient of interest). We used a Bonferroni correction of P < 4.3e−6 to correct for 11,628 pairwise tests.

Third, we tested for differences in the effects of males and females. We utilized a heterogeneity test (Cochran’s Q) to test for differences in the effect sizes, although this test is known to have low power. We first tested at previously and newly identified variants with a Bonferroni correction of P < 0.00032, and then expanded our search genome-wide (P < 5e−8).

Polygenic risk score and variance explained

For each individual in GERA and UKB, we constructed a polygenic risk score (PRS; weighted sum of the additive coding of each variant) utilizing 2 different coefficient estimates, and on 4 nested sets of variants. The increasing sets of variants included all variants from the previous set plus: (P) lead variants of loci identified previously; (PG) lead variants of loci identified in GERA; (PGC) conditionally independent SNPs identified in GERA. For the weights of each variant in the PRS, we used UKB coefficients when estimating variance explained in GERA, and GERA coefficients when estimating variance explained in UKB. For the novel SNPs discovered in the meta-analysis of GERA, UKB, and PAGE, there is no independent dataset to estimate beta coefficients, and so we do not report the variance explained for those variants. The variance explained was calculated by first computing the residual of QTc adjusted for age, sex, and genetic ancestry PCs. We then calculated the R² of the regression of that residual, adjusting for the PRS.

Heritability and family correlations

We first estimated familial heritability using twice the Pearson correlation for parent-offspring relationships and intraclass correlations (ICCs) for sib-pairs (2 × ICC) and second (4 × ICC) and third (8 × ICC) degree relatives (Visscher et al. 2008). We then estimated using more distant relatives by the additive array-based heritability of each individual’s long-term averaged covariate-adjusted QT_c in GERA using genome-wide complex trait analysis (GCTA) v1.93.2 (Yang, Lee, et al. 2011). Array heritability estimates require a large enough sample size (Lee et al. 2011), so we were restricted to non-Hispanic whites. We used only autosomal data, and LD filtered such that no 2 SNPs remained with pairwise r² > 0.8 (and conducted a sensitivity analysis with >0.3) with the PLINK v2.0 (Chang et al. 2015) greedy algorithm. Because of population stratification, as previously shown (Banda et al. 2015) we used PC-Relate (Conomos et al. 2016) rather than GCTA to estimate the kinship estimates, which does not make the assumption that the population is homogeneous (for a sensitivity analysis, we also estimate using standard GCTA kinship estimates). Using a greedy PLINK (Chang et al. 2015) removal algorithm, we filtered out individuals with kinship >0.025 (sensitivity analysis >0.05). Imputed data were included using best guess (as opposed to dosages), due to software limitations. We further compared results with LD score regression using LDSC v1.0.0 (Bulik-Sullivan et al. 2015), using LD score estimates using Europeans from the 1000 Genomes Project supplied by the software authors. Using LD score, we further tested for enrichment in overlapping categories of 10 tissue types and 24 main genetic categories as in previous work (Finucane et al. 2015) (enrichment defined as the ratio of the number of SNPs in high LD to the category vs the number of SNPs in low LD to the category).

Gain from multiple measures in GERA

We conducted several analyses to quantify the benefit of having multiple QT measurements on individuals. We first compared the effect sizes and PRS variance explained at each of a range of 1 through 5 QT measurements, and then all measurements. For this analysis, we used the 25,467 non-Hispanic white individuals who had at least 5 QT measurements available, to keep everything else identical in the comparisons. Second, we estimated the PRS variance explained due to, and in the absence of, measurement error, as has been described (Hoffmann et al. 2017).

Third, we estimated the effective sample size increase from using multiple measurements in GERA and estimated the equivalent sample size we would have needed if we had only 1 measurement. A general formula for the increase in sample size for the long-term average residual can be derived as follows. Suppose we have n individuals at k repeated measurements, and let Y_ij be the outcome for the ith individual at the jth measurement. Suppose that Y_ij follows a multivariate normal distribution with mean βX_i, where X_i is an individual’s genotype (the derivation does not depend on the allele frequency), with covariance matrix with 1 on the diagonal and the ICC ρ elsewhere. Here, ρ represents the within individual correlation of repeated measurements. Then the distribution of Y_i = k⁻¹∑Y_ij, the average of 1 individual’s phenotype (e.g. the long-term average residual), can be given by a normal distribution with variance σ² = k⁻¹[1 + (k − 1)ρ]. In a simple linear regression of Y_i = α + βX_i, we can see that the β estimate on average would not be affected by the number of measurements, but the SE, SE(β), given by σ[(n − 1)∑_i(X_i − mean(X_i))]^−1/2, will be. Now suppose that we wish to determine the effective sample size needed for k measures vs k^* measures. This is effectively determining n/n^* such that SE(β) = SE(β^*), which amounts to n/n^* = σ²(σ^*)⁻² = {k⁻¹[1 + (k − 1)ρ}/{(k^*)⁻¹[1 + (k^* − 1)ρ]}. For example, if ρ = 1 (completely correlated), the ratio is 1 (no benefit to having multiple measurements), and if ρ = 0, the ratio is k/k^* (e.g. if ρ = 0, and you have 5 measurements per person, that is equivalent to 5 times as many subjects).

Results

GERA

We conducted our primary discovery in the GERA cohort (n = 76,995; 81.8% self-reported non-Hispanic white, 8.1% Latino, 6.5% East Asian, 3.2% African American, 0.4% South Asian) (Table 1). The groups ranged from an average of 4.7–6.0 QT measurements per person. Females had higher QT_c measurements than males, consistent with prior work (Darpo et al. 2014). Non-Hispanic whites were estimated to have the highest QT_c, and South Asians the lowest. After correcting for age, sex, and medications (indicator variable each medication), compared with non-Hispanic whites, South Asians still had the lowest QT_c (β = −5.11: 95% CI = −7.63, −2.59), Latinos (β = −0.73; 95% CI = −1.31, −0.15), and African Americans (β = −0.71; 95% CI = −1.58, 0.171) were also estimated lower, and East Asians higher (β = 2.23; 95% CI = 1.57, 2.89). Of the 66 medications identified that we tested, 30 had P < 0.00076 (Bonferroni correction for 66 tests), and another 11 met P < 0.05 (Supplementary Table 1); the strongest association was with amiodarone (β = 19.3; 95% CI = 18.8, 19.8; P = 10⁻¹¹²²), known to have a strong effect on QT (Singh and Nademanee 1985).

Table 1.

Characteristics of GERA and the UKB.

	GERA NHW	GERA LAT	GERA EAS	GERA AFR	GERA SAS	UKB EUR
n (%)	62,976 (81.8%)	6,238 (8.1%)	5,034 (6.5%)	2,437 (3.2%)	310 (0.4%)	23,189 (100%)
Female, N (%)	36,530 (58.0%)	3,737 (59.9%)	2,869 (57.0%)	1,473 (60.4%)	115 (37.1%)	11,955 (51.6%)
Avg # QT meas (SD)	6.0 (7.4)	5.3 (6.8)	4.7 (6.2)	6.6 (8.6)	5.2 (7.6)	1
Avg # QT meas on med (SD)	3.9 (6.5)	3.4 (6.1)	2.5 (5.1)	4.4 (7.6)	2.9 (6.4)	0.2 (0.4)
Age at first meas (SD)	66.8 (11.8)	61.4 (13.3)	62.7 (12.7)	63.0 (12.4)	59.8 (12.3)	55.6 (7.5)
Female mean QTc (SD)	425.6 (24.4)	423.4 (23.5)	427.4 (24.2)	424.3 (26.5)	420.5 (23.4)	425.7 (26.0)
Male mean QTc (SD)	420.8 (26.5)	416.7 (25.5)	417.7 (25.3)	415.1 (26.6)	410.7 (23.4)	414.7 (25.1)

NHW, self-reported non-Hispanic whites; LAT, Latinos; EAS, East Asians; AFR, African Americans; SAS, South Asians; EUR, Europeans; meas, measurement; med, medication.

Variation in QT length by genetic ancestry

Within each self-reported race/ethnicity group, we tested for genetic ancestry PC correlations with QT_c. In non-Hispanic whites, PC₂ (P = 8.0e−5; variance explained a very modest r²_adjusted = 0.022%) was associated with QT_c; in East Asians, PC₁ (P = 0.047) and PC₂ (P = 0.011) were nominal (but would not meet a Bonferroni corrected significance level); all other groups showed no association in the interpretable first 2 PCs (Supplementary Table 2) (Banda et al. 2015). To better understand these relationships, we smoothed the phenotype distribution over the PCs (and labeled with ethnicity and nationality subgroups in Supplementary Fig. 1, see Materials and Methods).

Novel GERA loci and meta-analysis

We conducted a GWAS of QT_c in each GERA self-reported race/ethnicity group (adjusting for age, sex, and an indicator variable for each medication; Manhattan plots in Supplementary Fig. 2, a–e). Within each self-reported race/ethnicity group, non-Hispanic whites had 29 genome-wide significant loci), and Latinos had 3 genome-wide significant findings at variants that were previously reported, 2 of which were also genome-wide significant in East Asians, while African Americans and South Asians had no genome-wide significant loci, likely reflecting discovery sample size.

The meta-analysis of these 5 GERA groups (Supplementary Fig. 2f) had an inflation factor of λ = 1.064, which is very reasonable for a polygenic trait and a sample of this size (Yang, Weedon, et al. 2011). We identified a total of 32 genome-wide significant (P < 5e−8) loci, 9 of which were novel (Table 2), and 23 at previously reported loci. For these 9, upon replication in UKB + PAGE, all showed an effect in the same direction; 3 met a Bonferroni corrected P < 0.0056 and all but 1 met a nominal 2-sided P < 0.10 (this would be equivalent to a 1-sided P < 0.05 test, which is appropriate given the directionality requirement) testing the same direction of effect (suggesting these may be true associations as well). There was no evidence of heterogeneity at these novel variants among the 5 GERA groups (all had P > 0.05).

Table 2.

Lead variant at novel loci reaching genome-wide significance in GERA, with replication in the meta-analysis of UK Biobank (UKB) and PAGE.

SNP	Gene	b37 Chr	b37 Pos	A1	A2	Freq A1 GERA	Eff A1 GERA	P GERA	N GERA	Eff A1 UKB + PAGE	P UKB + PAGE	N UKB + PAGE
rs12476527	KCNK3O	2	26915624	G	T	0.590	−0.66	3.1e−10	76,995	−0.27	0.083	40,537
2:37210516_CAAA_C	HEATR5BCI	2	37210516	CAAA	C	0.487	0.61	4.2e−09	76,995	0.35	0.017	40,537
rs59212267	THRBHI	3	24464955	G	A	0.300	0.60	4.7e−08	76,995	0.32	0.053	40,537
rs6775742	MITFC	3	69773998	A	G	0.316	−0.61	2.3e−08	76,995	−0.47	0.0026	40,537
rs56052942	ATP1B3C	3	141649607	T	A	0.106	0.92	2.5e−08	76,995	0.44	0.084	40,537
rs6839228	CAMK2DHI	4	114444701	G	C	0.238	0.65	4.4e−08	76,995	0.24	0.33	23,189
rs2017201	RP11-772C9.1C	5	56999693	A	G	0.821	0.90	1e−11	76,995	0.87	1.8e−05	40,537
rs8024538	ALPK3CI	15	85368419	A	T	0.562	0.63	1.7e−09	76,995	0.54	0.00044	40,537
rs8133749	LINC00189/BACH1AI	21	30615935	C	T	0.490	−0.85	4.8e−17	76,995	−0.39	0.0075	40,537

Coded to the A1 allele. H, heart eQTL; A, Artery eQTL; O, other eQTL; C, closest gene; 3, 3'-UTR; I, intron; M, missense.

In our meta-analysis of GERA, UKB, and PAGE, we identified an additional 19 novel loci (Table 3 and Supplementary Fig. 2g; λ = 1.053). All had the same direction of effect in all 3 cohorts.

Table 3.

Lead variant at novel loci reaching genome-wide significance in the meta-analysis of GERA, UKB, and PAGE.

SNP	Gene	b37 Chr	b37 Pos	A1	A2	Freq A1	Eff A1	P	N	Sign
rs706007	CASZ1CI	1	10780727	A	G	0.115	0.96	1e−10	117,532	+++
rs608930	RP1-79C4.1/PRRX1H	1	170617306	G	T	0.584	−0.47	4.8e−08	117,532	−−−
2:169008248_CA_C	STK39OI	2	169008248	CA	C	0.342	0.56	6.5e−09	117,532	+++
rs78598566	AC098617.2C	2	192723897	C	T	0.100	0.87	3.3e−10	117,532	+++
rs259542	ZNF385DCI	3	21954382	T	A	0.409	−0.50	7.6e−09	117,532	−−−
rs574752230	SLC25A20C	3	48943646	C	CA	0.554	0.83	8.4e−12	117,532	+++
rs748142903	SLC27A6CI	5	128307280	TTTG	T	0.220	0.62	1.5e−08	117,532	+++
rs74786782	CDK13CI	7	40123256	A	G	0.127	−0.68	4.6e−08	117,532	−−−
rs10957135	RP11-163N6.2H	8	61309452	C	A	0.292	0.51	3e−08	117,532	+++
rs72671655	ZFPM2CI	8	106347897	T	A	0.034	1.28	4.6e−08	117,222	+++
rs7865239	C9orf3H	9	97723449	C	T	0.402	0.50	4.1e−09	117,532	+++
rs7098978	ARMC3H	10	23144216	C	G	0.330	−0.55	7.6e−10	117,532	−−−
rs3878005	PIP4KAH	10	75431078	C	A	0.148	0.67	3.4e−08	117,532	+++
rs747782	C1QTNF4/SLC39A13/MADDH	11	47940925	T	C	0.202	0.63	1.1e−09	117,532	+++
11:65263818_CAG_C	MALAT1C	11	65263818	CAG	C	0.115	−0.81	1.6e−08	117,532	−−−
rs11874	NSFP1A3	17	45017193	G	A	0.105	0.87	2.1e−10	117,532	+++
rs2287223	HLFOI	17	53392852	G	A	0.419	−0.49	1.7e−08	117,532	−−−
rs8067445	MXRA7HI	17	74687777	G	T	0.417	0.47	3.3e−08	117,532	+++
rs661821	ZNF358OI	19	7582649	T	C	0.632	0.59	1.7e−09	117,532	+++

Coded to the A1 allele. H, heart eQTL; A, Artery eQTL; O, other eQTL; C, closest gene; 3, 3'-UTR; I, intron; M, missense.

Sensitivity analysis including vs not including medication

To assess what impact ignoring medication would have on a GWAS, we compared results of a GWAS of treated QT measurements vs a GWAS of untreated in GERA (see Materials and Methods). We saw little difference in the overall significance patterns in a Chicago plot comparing the 2 GWAS (Supplementary Fig. 3) and also when comparing the effect sizes at previously identified variants (Supplementary Fig. 4). A formal test for the difference between treated and untreated measurement SNP effects showed no differences genome-wide (λ = 1.024, Supplementary Fig. 5); focusing on all previously identified and novel QT interval variants, no variant met P < 0.00032 (correction for 153 tests); and further focusing on the 54 previous pharmacogenetic results, none met P < 0.00093 (correction for 54 tests).

Replication of previously identified SNPs

A total of 31 of 36 loci previously associated as genome-wide significant with QT length replicated (P ≤ 0.0014) in a meta-analysis of GERA and the UKB, and all but 1 variant met nominal P ≤ 0.05; for previously reported variants with P > 5e−8, the replication rate was much lower, with 6/74 replicating (P ≤ 0.00068), and an additional 5 meeting a nominal P < 0.05, all in the same direction. Full details are in Supplementary Table 3. Only 1 variant, the previously reported variant rs7122937, showed ethnic heterogeneity (I² = 80.2, P_{heterogeneity} = 0.00045; β_{non-Hispanic white} = 2.38, 95% CI = 2.09–2.68; β_Latino = 1.03, 95% CI = 0.29–1.77; β_{East Asian} = 1.06, 95% CI = 0.21–1.90; β_{African American} = 1.36, 95% CI = (0.23, 2.49); β_SAS = 0.37, 95% CI = −2.39 to 3.13), all other variants had P_{heterogeneity} ≥ 0.005.

Conditional analysis

A strength of a large single cohort is more accurate conditional analysis; we next sought to find additional independent signals in GERA. A total of 9 loci had a total of 15 additional independent variants. The vast majority of these had very weak LD with the original lead variant; 11 of these 15 replicated (P < 0.003) in the UKB, and an additional 2 met a nominal P < 0.05 (Supplementary Table 4). We then conditioned on all previously reported variants to determine if these were novel independent associations; 3 variants remained genome-wide significant: the less common rs56128851 (NOS1AP; MAF 0.01; β = −3.95, SE = 0.53, P = 9.5e−14; r² ≤ 0.006 with all previously reported variants), and common variants rs57248046 (DPT, MAF 0.053, β = 1.44, SE = 0.25, P = 4.7e−9, r² ≤ 0.12 with all previously reported variants) and the indel 2:179768491_TA_T (CCDC141, MAF 0.142, β = 0.90, SE = 0.15, P = 1.8e−9, r² ≤ 0.037 with all previously reported variants).

Characterizing dominance, epistasis, and sex differences

We next sought to examine the QT-associated variants for dominance and epistatic effects. We found no evidence of dominance at individual variants (P < 0.00032, Bonferroni correction for all 153 previously and newly identified independent variants: 110 previously reported, 9 GERA discovery, 19 GERA, UKB, PAGE meta-analysis discovery, and 15 conditional), and no evidence for epistasis (P < 4.3e−6, Bonferroni correction for all pairwise tests).

We further ran GWAS in GERA stratified by sex, but no additional loci were found (Supplementary Fig. 6), as might be expected due to reduced sample size, nor were any variants genome-wide significantly different between the sexes (although we expected power to be very low). However, one of the previously identified and replicated variants met a Bonferroni correction (P < 0.00032). The previously reported variant rs12143842 (NOS1AP, the variant with the strongest association with QT) differed between the 2 groups with a stronger effect in females (P = 5.5e−6; β_male = 3.09, P_male = 6.2e−50; β_female = 4.30, P_female = 4.0e−150), a difference previously identified (Winbo et al. 2017).

Variance explained and conditional SNP impact

We assessed the proportion of QT variance explained using a PRS of previously and newly identified variants (Table 4). We first looked at variance explained in UKB whites. An estimate of variance explained using the lead variant at previously identified loci (P, see Materials and Methods) is estimated at 3.9% (95% CI 3.4–4.3%). Including lead variants at GERA-identified loci (PG) did not improve this estimate. When we additionally include conditionally independent variants at all previous loci (PGC), variance increased to 5.9% (95% CI 5.3–6.5%). Results were similar in GERA non-Hispanic whites, estimated for P at 3.5% (95% CI 3.2–3.8%), a slight increase for PG at 3.8% (3.5–4.1%), and a total of 5.5% (95% CI 5.2–5.8%) for PGC. When we compare the variance explained by the PRS among the self-reported race/ethnicity groups within GERA for the PGC set, we see that most groups have very overlapping CIs with Europeans, with the exception of African Americans, whose variance explained is substantially less at 0.5% (95% CI 0.1–1.3%); the lower variance explained in African Americans has been seen before for other traits. We did not include variants identified in the meta-analysis of GERA, UKB, and PAGE, since no independent cohort or coefficients were available.

Table 4.

Variance explained [P-value] from PRS in each group.

Group	Coefficient	P P	P PG	P PGC
GERA NHW	UKB	0.035 (0.032, 0.038) [1e−496]	0.038 (0.035, 0.041) [1e−538]	0.055 (0.052, 0.058) [1e−792]
GERA LAT	UKB	0.039 (0.030, 0.049) [1e−55]	0.043 (0.034, 0.052) [5.2e−61]	0.051 (0.041, 0.062) [4.4e−73]
GERA EAS	UKB	0.034 (0.026, 0.044) [4.9e−40]	0.036 (0.026, 0.047) [1.8e−42]	0.040 (0.031, 0.052) [4.9e−47]
GERA AFR	UKB	0.005 (0.001, 0.013) [0.00028]	0.006 (0.001, 0.015) [7.9e−05]	0.021 (0.011, 0.035) [8.0e−13]
GERA SAS	UKB	0.057 (0.017, 0.116) [2.3e−05]	0.072 (0.027, 0.133) [1.7e−06]	0.060 (0.016, 0.118) [1.3e−05]
UKB White	GERA	0.039 (0.034, 0.043) [1e−203]	0.038 (0.034, 0.044) [1e−203]	0.059 (0.053, 0.065) [1e−316]

P, previously identified; PG, P, plus GERA; PGC, P, G, plus additional conditionally independent variants at loci; NHW, non-Hispanic whites; LAT, Latinos; EAS, East Asians; AFR, African Americans; SAS, South Asians; EUR, Europeans.

Heritability from all SNPs and family correlations

Family heritability estimates were 28.4% for parent–child (95% CI = 19.7%, 37.1%; broken down by sex in Supplementary Table 5), 41.4% for siblings (95% CI = 30.2%, 52.4%), 27.1% for second degree relatives (95% CI = 2.9%, 51.1%), and 43.0% for third degree relatives (95% CI = 7.7%, 78.3%); combining these gives a heritability estimate of 30.9% (95% CI = 26.2%, 35.5%). Array-based heritability estimated in non-Hispanic white GERA individuals based on imputed genotypes using the PC-Relate kinship estimate (adjusted for population stratification) was 17.1% (95% CI = 15.5%, 18.7%); the estimate was slightly higher when using the GCTA kinship estimate (24.5%; 95% CI = 22.1%, 27.0%). The LD score estimate was similar (16.8%; 95% CI = 11.4%, 22.1%). With LD score, we also partitioned the heritability and tested for enrichment of 10 tissue types and 24 main genetic categories [see Materials and Methods; Finucane et al. (2015)]; cardiovascular tissue had the highest and most significant estimate of enrichment (7.2, 95% CI = 5.7–8.7, P = 2.1e−10, Fig. 2) for tissue type, and super enhancer was the most statistically significant enriched region in the genome although not the strongest estimated enrichment (2.96, 95% CI = 2.26–3.66, P = 3.1e−7; Fig. 2). Sensitivity analyses are included in Supplementary Table 6.

Fig. 2.

Partitioned heritability enrichment. Heritability enrichment estimates in 10 cell types (a), with significance (b), and for 24 genomic categories (c), with significance (d).

Gain from multiple measurements

We tested the benefit of having multiple QT measurements in an analysis restricted to individuals with ≥5 measurements, using a PRS of all previously and newly identified lead variants plus conditional independent variants. Using all measurements, as opposed to just one, reduced the SE of the regression coefficient by 24% and increased the variance explained by 163% (Supplementary Table 7). We also estimated the equivalent sample size for a study of single observations per individual compared with our sample which had 6 measurements per person. We found that our 80,000 individuals (approximate GERA sample size) with 6 measurements per person and an ICC of 0.30 (covariate-adjusted ICC from GERA), would be equivalent to a sample 6/(1 + 5 × 0.3) = 2.4 times larger, or 192,000 individuals, with a single measurement.

Discussion

In the large, ethnically diverse GERA cohort, we identified genetic ancestry PC associations with QT interval within non-Hispanic whites and East Asians; 9 novel independent loci in GERA, most replicating at nominal significance; 19 novel loci in the meta-analysis of GERA, the UKB, and PAGE; 3 novel additional independent variants at known loci; a substantial increase in variance explained from the novel loci, including the novel conditional variants; and overall array heritability ≈17%, which was about half of the family based heritability of 31%. In addition, including multiple measurements increased the variance explained by 163% and was equivalent to a study using a single measurement of 2.4 times the sample size. We also noted that previous QT GWAS results were likely not influenced by failure to adjust for medications, as we failed to see any difference in association of putative QT interval variants with treated (grouped across medications) vs untreated individuals.

There was biological support in the aggregate; partitioning across different tissue types, we noted enrichment in cardiovascular tissue across the genome (enrichment 7.1, 95% CI = 5.7–8.7). There was also biological support for the novel loci identified; of the loci identified in GERA and the meta-analysis, 9 SNPs were heart tissue eQTLs, 2 in atrial tissue, and 4 additional were other tissue eQTLs in GTEx. Almost all variants identified had MAF >5%, except the conditional variant rs56128851 in NOS1AP (1.0%).

Many of our variants have been previously reported in association with cardiovascular related phenotypes. Of those identified in the GERA analysis, variants in LD (r² ≥ 0.8) with our lead variant in KCNK3 have been previously reported associated with diastolic, systolic, and pulse blood pressure (Hoffmann et al. 2017); mean arterial pressure (Kato et al. 2015), CVD (Kichaev et al. 2019), and lipid levels (Ehret et al. 2016). Our lead variants in KCNK3, HEATR5B, THRB, ATP1B3, CAMK2D, and ALPK3, have been associated with the PR interval (which precedes the QT interval) (Ntalla et al. 2020) and atrial fibrillation (Roselli et al. 2018); of particular note, this included all of GERA identified loci that failed to replicate (P > 0.05) in the meta-analysis of UKB and PAGE. In addition, our lead variant at RP11-772C9.1 is in LD with variants previously associated with hypertropic cardiomyopathy (Tadros et al. 2021).

Of those identified in our meta-analysis of GERA, UKB, and PAGE, variants in LD (r² ≥ 0.8) with our lead variant in STK39 were associated with atrial fibrillation (Ellinor et al. 2012), early-onset atrial fibrillation (Lee et al. 2017), ischemic stroke (Malik et al. 2018), and arrhythmia (Ishigaki et al. 2020); AC096817.2, STK39, ZFPM2, and ZNF358 had also been found to be associated with the PR interval (Ntalla et al. 2020); C9orf3 and SYNPO2L with atrial fibrillation (Roselli et al. 2018); PIP4KA with SBP (Giri et al. 2019; Kichaev et al. 2019); PIP4KA, MALAT1, ZNF358, and ZFPM2 with ECG amplitude (Verweij et al. 2020); NSFP1 with blood pressure traits (Hoffmann et al. 2017), coronary artery disease (van der Harst and Verweij 2018), QRS duration (Evans et al. 2016; Prins et al. 2018); MXRA7 with pulse pressure (Evangelou et al. 2018; Giri et al. 2019); and ZNF358 and NSFP1 with myocardial fractal dimension (Meyer et al. 2020). In addition, mutations in the CASZ1 gene have been associated with cardiomyopathy (Qiu et al. 2017, p. 1) and all-cause mortality in a cardiovascular cohort (Abdulrahim et al. 2019). The variant rs6749447 in STK39 has been associated with cardiovascular complications (r² = 0.7 with 2:169008248_CA_C, identified here) (Kunnas et al. 2021). Variants in SLC25A20 have been associated with carnitine–acylcarnitine translocase deficiency which can result in arrhythmias and cardiac arrest (Rubio-Gozalbo et al. 2004). Variants in SLC27A6 have been associated with left ventricular hypertrophy (Auinger et al. 2012). In summary, it appears that QT interval has a multifactorial genetic etiology, involving different aspects of cardiac function.

Our array heritability estimates of ≈17% with both PC-relate kinship estimates through GCTA, and with LD score, are consistent with previous estimates (Yang, Manolio, et al. 2011). Previous studies have found similar or slightly higher values for variance explained by identified SNPs (range from ≈5% to 10%), whereas the highest variance explained we found in GERA whites using UKB weights was 5.5% using multiple measurements, and variance explained we found in UKB whites using GERA weights was 5.9%. We suspect this might have something to do with cohort differences, and prior studies potentially using in sample beta coefficients. We found differences in variance explained by self-reported race/ethnicity, where East Asians had a somewhat lower value and African Americans a much lower value, while the value for Latinos was comparable to whites. These differences are similar to those found for other traits and likely reflects SNP discovery in cohorts of largely European genetic ancestry.

One limitation of our comparison of individuals on and off drugs is that it included multiple different drugs which were putatively QT increasing. Nevertheless, we saw little difference in estimated effects between individuals on and off drugs, and our GWAS analysis for GERA and UKB included an adjustment for each drug if present at every individual measurement in the dataset. Future work entails assessing pharmacogenetic effects of each of these drugs individually. Another limitation was that PAGE did not adjust for the same medications as was done in GERA and UKB, and may have added heterogeneity, leading to diminished discovery power. However, given the modest difference in estimated effects between individuals on and off drugs in GERA, we expect the reduction in power to be small.

In summary, our analysis of extensive longitudinal data in the GERA cohort, which provided multiple measurements of QT interval per subject, enabled identification of novel QT interval variants and increased understanding of genetic contributions to QT interval. By our estimates, QT interval has a moderate heritability. It is an important clinical trait because lengthening of the interval can lead to adverse cardiac outcomes such as cardiac arrest and sudden death. Expanding the genetic variance explained for this parameter has the potential to provide predictive value in the clinical setting. Further exploration of these genetic variants in association with response to drugs that lengthen QT may be particularly useful in the clinical setting.

Data Availability

Summary statistics of all analyzed markers and cohorts (estimates, SE, and N) are available from the NHGRI-EBI GWAS Catalog (https://www.ebi.ac.uk/gwas/downloads/summary-statistics) under study accession numbers GCST90165290 and GCST90165291. To protect individual’s privacy, complete GERA data are available upon approved applications to the Kaiser Permanente Research Bank Portal (https://researchbank.kaiserpermanente.org/for-researchers), and complete UKB data are available upon approved applications to the UK Biobank (https://www.ukbiobank.ac.uk). PAGE summary statistics were obtained from the GWAS catalog under study accession number GCST008043. A detailed list of QT interval drugs can be obtained upon registration from https://www.crediblemeds.org. GTEx data were obtained from the GTEx portal (www.gtexportal.org) and can be obtained from dbGaP accession number phs000424.v8.p2.

Supplemental material is available at GENETICS online.