Multi-ethnic Genome-wide Association Study of Decomposed Cardioelectric Phenotypes Illustrates Strategies to Identify and Characterize Evidence of Shared Genetic Effects for Complex Traits

Abstract

Background -

We examined how expanding electrocardiographic (ECG) trait genome-wide association studies (GWAS) to include ancestrally diverse populations, prioritize more precise phenotypic measures, and evaluate evidence for shared genetic effects enabled the detection and characterization of loci.

Methods -

We decomposed 10-second, 12-lead ECGs from 34,668 multiethnic participants (15% African American; 30% Hispanic/Latino) into six contiguous, physiologically-distinct (P wave, PR segment, QRS interval, ST segment, T wave, and TP segment) and two composite, conventional (PR interval and QT interval) interval-scale traits and conducted multivariable-adjusted, trait-specific univariate GWAS using 1000-G imputed SNPs. Evidence of shared genetic effects was evaluated by aggregating meta-analyzed univariate results across the six continuous ECG traits using the combined phenotype adaptive sum of powered scores test (aSPU).

Results -

We identified six novel (CD36, PITX2, EMB, ZNF592, YPEL2, and BC043580) and 87 known loci (aSPU p-value<5E-9). Lead SNP rs3211938 at CD36 was common in African Americans (minor allele frequency=10%), near-monomorphic in European Americans, and had effects on the QT interval and TP segment that ranked among the largest reported to date for common variants. The other five novel loci were observed when evaluating the contiguous, but not the composite ECG traits. Combined phenotype testing did not identify novel ECG loci unapparent using traditional univariate approaches, although this approach did assist with the characterization of known loci.

Conclusions -

Despite including one-third as many participants as published ECG trait GWAS, our study identified six novel loci, emphasizing the importance of ancestral diversity and phenotype resolution in this era of ever-growing GWAS.

Introduction

Genetic susceptibility underlies a majority of common diseases and traits, as demonstrated by genome-wide association studies (GWAS) that have identified thousands of genetic loci for cardiovascular, cardiometabolic, cancer, kidney, psychiatric, ocular, inflammatory, and neuromuscular traits¹. Together, these GWAS have revealed common threads underlying the genetic architecture of complex diseases and traits, as well as research gaps. For example, evidence of shared genetic effects (i.e., pleiotropy) is widespread, even for traits with few known etiologic links^2,3. Yet few studies have systematically examined evidence of shared genetic effects, thereby missing opportunities to identify and characterize master regulators as strong candidates for intervention^2,4. There is also limited racial/ethnic diversity in published GWAS, as the majority (>80%) of GWAS have been conducted in European ancestral populations³. Limited diversity leads to a biased view of human variation that hinders translation of genetic associations into clinical and public health applications for all populations^5,6. Further, the scale and collaborative nature of GWAS prioritize traits that are widely available, although these traits may not precisely capture phenotypic variation and underlying biology^7,8. Together, these research gaps argue for expanding GWAS analyses to systematically examine evidence of shared genetic effects across a spectrum of biologically motivated traits in multi-ethnic populations.

Electrocardiograms (ECG) measure a sequence of distinct electrophysiological processes in the myocardium that underlie cardiac conduction and repolarization. ECG traits have high heritability⁹, are relevant to cardiovascular health¹⁰, and allow opportunities for dense phenotyping¹¹. Moreover, there are few racially/ethnically diverse GWAS of ECG traits¹². Therefore, ECG traits are well suited for assessing the degree to which increased racial/ethnic diversity, evaluation of genetic effects shared across phenotypes, and improved phenotype resolution can enhance locus identification and characterization. We therefore examined individual and shared genetic effects underlying six contiguous measures of the ECG waveform spanning an average heartbeat using data from the multi-ethnic Population Architecture using Genetic Epidemiology (PAGE) study and the Multi-Ethnic Study of Atherosclerosis (MESA). Our results illustrate the broad utility of multi-ethnic GWAS of carefully constructed individual and aggregate traits to illuminate the biology of complex diseases and traits.

Methods

Methods for this paper are detailed in online supplementary material.

Summary-level (PAGE) and individual-level (ARIC, HCHS/SOL, MESA, and PAGE) data are available at DbGaP (https://www.ncbi.nlm.nih.gov/gap/, accession numbers phs000090.v1.p1 [ARIC], phs000810.v1.p1 [HCHS/SOL], phs000293.v1 [MESA], phs000.56.v1.p1 [PAGE], and phs000200.v1 [WHI]).

The institutional review board of the University of North Carolina at Chapel Hill determined this study as exempt from review, further each participating study was approved by the institutional review board at the respective sites, and all participants provided written consent.

Results

Sample description

Of the 39,538 participants with GWAS and ECG data in ARIC, HCHS/SOL, MESA, and WHI, 34,668 (88%) met all inclusion criteria (Table S1, Table S2). Seventy-five percent of eligible participants were female, the mean age was 55 years, and nearly half were either Hispanic/Latino (30%) or African American (15%) (Table S3). On average, participants were overweight (BMI mean = 29 kg/m²) and had high serum low-density lipoprotein cholesterol (mean = 135 mg/dL). There was a high prevalence of hypertension (49%). Holding all adjustment variables constant, PR segment and TP segment durations were the most strongly correlated among the six ECG traits (partial correlation ρ=−0.64), whereas T wave and P wave durations (ρ=−0.01) was largely uncorrelated (Table S4). P wave and QRS interval were the only two ECG traits with significant and positive genetic correlations (r_g=0.27; p=0.05) (Table S5).

Overview of association results

Approximately 22M SNPs met our inclusion criteria (Table S6) and were evaluated in our combined phenotype multi-ethnic analysis of six contiguous ECG traits, our primary analysis (Figures S1, S2). Lead SNPs at 82 of 149 loci (56%) previously reported by 26 interval scale ECG trait GWAS analyses (Table S7) were identified at genome-wide significance in our multi-ethnic population. The identification of known loci varied by trait (Table S8), ranging from 21 of 45 (47%) loci for QRS interval to nine of 14 (64%) loci for P wave. When using a lower significance threshold of p_aSPU<0.0003 (0.05/149), 123 of the 149 (83%) previously recognized interval scale ECG trait loci were identified.

An additional six loci identified by our primary analysis were >2 Mb away from all lead SNPs previously reported by interval scale ECG trait GWAS and are presented as novel (Table 1, Figure 2). As described below, our results highlight the utility of phenotype decomposition, ancestral diversity, and combined-phenotype testing for the identification and characterization of complex trait loci.

Table 1.

Novel genome wide-significant (paSPU < 5×10–9) loci discovered in genome-wide association study of six contiguous electrocardiographic traits that decompose an average heartbeat in N=34,668 participants from the multi-ethnic Population Architectures using Genomics and Epidemiology study (PAGE) and the Multi-Ethnic Study of Atherosclerosis (MESA)

			Coded allele	Non- coded allele						Contiguous ECG traits						Composite ECG traits
Lead SNP	Cnr	Position			Locus	AA	EA	CHN	HIS	P wave	PR segment	QRS interval	ST segment	T wave	TP segment	QT interval	PR interval
rs13143308	4	111714419	T	G	PITX2	30%	21%	74%	39%	2×10−11	5×10−4	2×10−1	3×10−2	3×10−1	1×10−1	4×10−1	8×10−1
rs4340921	5	49687697	C	T	EMB	66%	46%	49%	44%	8×10−13	8×10−1	1×10−3	7×10−1	2×10−1	2×10−4	7×10−1	2×10−3
rs3211938	7	80300449	G	T	CD36	10%	<0.01%	<0.01%	1%	2×10−5	8×10−3	5×10−3	4×10−1	1×10−5	1×10−13	6×10−10	6×10−6
rs11073663	15	85260268	A	G	ZNF592	27%	54%	19%	48%	4×10−1	3×10−10	5×10−1	4×10−1	2×10−3	7×10−3	2×10−2	6×10−7
rs142166837	17	57471022	C	T	YPEL2	31%	52%	32%	49%	5×10−2	6×10−1	1×10−3	1×10−1	4×10−11	1×10−2	4×10−7	8×10−1
rs13047360	21	28851580	G	A	BC043580	7%	17%	23%	16%	7×10−1	3×10−2	2×10−11	2×10−1	5×10−2	1×100	2×10−1	3×10−2

aSPU: adaptive sum of powered tests

AA: African American, CHN: Chinese-American, EA: European-American, HIS: Hispanic/Latinos

Bolded values exceed the genome-wide significance threshold (p<5×10⁻⁹)

Figure 2.

Lead SNPs at 87 loci significantly associated (p_aspu<5×10⁻⁹) with six contiguous ECG traits spanning an average heartbeat, in n=34,668 multi-ethnic participants in the Population Architecture Using Genomics and Epidemiology study and Multi-Ethnic Study of Atherosclerosis. Outer stars denote novel loci and darker shades of green indicate lower p-values. To aid interpretation, lead SNPs were organized into broadly similar groups using hierarchical cluster analysis.

Phenotype decomposition

Of the six novel loci identified in our primary multi-ethnic combined phenotype analysis, accompanying univariate analyses indicated that lead SNPs primarily affected P wave (PITX2, EMB), TP segment (CD36), PR segment (ZNF592), T wave (YPEL2), and QRS interval (BC043580). None of the novel loci were associated with ST segment. Further, the combined phenotype analysis did not identify novel loci beyond univariate analysis.

We then contrasted results for the six contiguous ECG traits with results from the two composite ECG traits, QT interval (QRS interval + ST segment + T wave) and PR interval (P wave + PR segment) (Figure S3). CD36 was the only novel locus identified for both a contiguous (TP segment) and a composite (QT interval) ECG trait (Table 1). We also examined evidence of consistency of SNP effects by grouping traits according to whether they affected atrial (PR interval, PR segment, and P wave) or ventricular (QT interval, QRS interval, T wave, and ST segment) conduction. For atrial traits, novel loci identified for the contiguous traits had varying directions of effects (Figure S3a, Table S9), which when combined resulted in near-zero estimated effects for the composite trait. For example, every copy of the T allele for PITX2 lead SNP rs13143308 increased P wave duration by 0.63 milliseconds [ms] (p_univariate=2×10⁻¹¹), but shortened the PR segment by 0.58 ms (p_univariate=6×10⁻⁴). However, when evaluated together as the composite trait PR interval, every copy of the rs13143308 T allele prolonged the PR interval by 0.03 ms (p_univariate=0.84). Similarly, among the 59 loci associated with ventricular conduction, two of the three novel loci (rs142166837 and rs13047360) had opposite effects on QRS interval and T wave duration, which did not reach genome-wide significance when summed for the composite trait QT interval (Figure S3b, Table S9). There were no instances of either PR or QT interval identifying a novel locus not associated with any of the six contiguous traits at the genome-wide level.

Ancestral diversity

Lead SNPs at five of the six novel loci were common (MAF >5%) across ancestral populations, with modest evidence of heterogeneity of effect across race/ethnicity (Table S10). One locus (CD36) showed evidence of population specificity, with lead SNP rs3211938 near-monomorphic in European and Chinese populations (MAF<0.01%), infrequent in Hispanic/Latinos (MAF=1%), and common in African Americans (MAF=10%). Variant rs3211938 showed genome-wide significant associations with TP segment (p_univariate=1×10⁻¹³) and QT interval (p_univariate=6×10⁻¹⁰) and nominal associations with P wave (p_univariate=2×10⁻⁵), PR segment (p_univariate=0.008), and QRS interval (p_univariate=0.005). Although no GWAS of TP segment has been published, each copy of the rs3211938 G allele increased QT interval by 3.70 ms. Reported effects for common (MAF>5%) QT lead SNPs range from 0.5 ms to 3.5 ms¹³. SNP rs3211938 was either genotyped or well-imputed across studies and ancestry groups (imputation quality > 0.98, Table S11).

Combined phenotype analyses

We found widespread evidence of shared genetic effects across ECG traits, with aSPU gamma scores that varied substantially across lead SNPs (Table S9). One fourth of lead SNPs identified as genome-wide significant (p_aspu<5×10⁻⁹) had univariate associations with at least two ECG traits (p_Univariate<5×10⁻⁹). Lead SNPs at ACVR2B, SCN5A, SCN10A, CEP85L, CAV1, and TBX5 were associated with three or more ECG traits at univariate genome-wide significance levels. As expected, traits that were more highly correlated also showed stronger evidence of shared genetic effects, with 10 of the 20 lead SNPs that were associated with PR segment also showing genome-wide associations with TP segment. However, evidence of shared genetic effects among uncorrelated traits was also observed. For example, eight genome-wide significant SNPs at SCN5A, SCN10A, CEP85L, and CNOT1 exhibited significant univariate associations with both the T wave and P wave, despite low correlation between the two traits (ρ = −0.01).

There also was evidence of allelic heterogeneity for multiple ECG traits. As an example, five signals in low LD (r² < 0.1) were detected within a 500 kb region near the previously identified locus TBX5, each associated with a distinct combination of ECG traits. Lead variants at these five independent signals remained genome-wide significant after sequential conditional analyses (results not shown). Two of the five independent signals (rs3741698 and rs2047752) were largely specific to PR segment (Figure 3). The other three signals involved PR segment and QRS interval (lead SNP rs4784657), P wave, QRS interval, and TP segment (lead SNP rs883079), and the combined phenotype only (lead SNP rs1895595). Lead SNPs also showed some evidence of variation across traits, including the locus identified by lead SNP rs7484657, for which p-values for the QRS interval lead SNP differed by approximately three orders of magnitude from the rs7484657 p-value.

Figure 3.

Regional SNP associations and linkage disequilibrium at four independent signals near TBX5 among 34,668 participants with electrocardiographic data in the Population Architectures using Genomics and Epidemiology (PAGE) study and Multi-Ethnic Study of Atherosclerosis. Lighter colors indicate greater linkage disequilibrium with lead SNPs, and black markers denote SNPs not in LD (r² < 0.1) with any of the four lead SNPs. Combined phenotype p-values are truncated at 1×10⁻¹¹.

Bioinformatic characterization

The rs3211938 variant is a well-known, non-synonymous protein coding variant causing CD36 deficiency¹⁴. For the five remaining novel SNPs, bioinformatics characterization found evidence of genetic regulation, including chromatin marks, and regulatory motifs (Table S12). Each of the novel lead SNPs were either rated as evolutionarily conserved based on the GRASP conservation score, or were in high linkage disequilibrium with another SNP meeting that threshold (R² > 0.9). In addition, rs1107366 was identified to regulate the expression of several long non-coding RNAs in ECG-relevant tissues (adipose, arterial, atrial tissue).

Discussion

In this study, we examined the extent to which combined multi-ethnic GWAS analyses of carefully selected phenotypes that map to well-defined biology improved detection and characterization of ECG trait loci. We identified six novel loci, five of which were detected only when examining the more precisely defined phenotypes, and a sixth locus that was specific to African ancestral populations. We also showed how leveraging evidence of a shared genetic architecture aided the characterization of known loci, particularly when loci harbored multiple independent signals that differed by trait. In this mega-GWAS era involving predominantly European ancestral populations, this study, conducted in a population one-third the size of the largest published ECG trait GWAS^13,15, underscores the merits of prioritizing diversity and phenotype measurement.

Of the three GWAS challenges we examined, our deliberate selection of phenotype measures mapping to well-defined biology largely drove locus discovery, challenging current trends in GWAS that emphasize increased sample size. The growing scale of GWAS, which today can surpass one million participants¹⁶, has resulted in the prioritization of commonly available traits (e.g., body mass index) over traits that more precisely capture underlying biology (e.g., direct measures of body fat¹⁷). In our case, composite ECG traits PR interval and QT interval have been most commonly interrogated by GWAS. However, these traits represent aggregates of physiologically distinct mechanisms, which may obscure loci with effects localized to, or inconsistent across, individual contiguous traits. This phenomenon was illustrated by the PITX2 locus, a locus associated with atrial fibrillation¹⁸. Because PITX2 lead SNP rs13143308 had opposing associations with the contiguous P wave and PR segment, a standard approach using the composite PR interval yielded a near-zero effect, obscuring the potential importance of the locus on atrial function regardless of sample size. These results emphasize the need to balance ongoing investments in large-scale genome measurement with use of precision phenotyping, for instance through efforts like the ongoing Precision Medicine Initiative’s All of Us Research Program¹⁹.

The six traits we used in our ECG decomposition were motivated by their relations to physiology, and their coherence as an aggregate electrophysiologic phenotype. While an important complement to traditional, coarser measures like the PR and QT intervals, our phenotype decomposition approach that identified novel loci and improved characterization of known loci captured but a fraction of the full variation in ECG phenotypes. Further phenotypic specificity and additional biologic insight may be offered by GWAS of other ECG traits, including measures of waveform amplitudes, angles, or variability. For example, a recent, as yet unpublished of UK Biobank data used each of the sampled amplitudes recorded on the digital ECG, forming hundreds of distinct ECG measures for separate evaluation in GWAS (https://www.biorxiv.org/content/10.1101/648527v1). Another approach might focus on traits governed by a plausibly-shared genetic architecture, such as ion channel function or cardiac remodeling, potentially assisting efforts to map loci to specific biologic pathways. Further extending combined phenotype ECG trait GWAS to include other phenotypes and traits (e.g. cardiometabolic traits or cardiovascular diseases) also is warranted, given evidence that these traits represent interrelated manifestations of common biologic mechanisms¹² and the success of prior combined phenotype studies to disentangle complex biology²⁰. Overall, how to select intermediate traits and integrate such traits with other phenotypic data, including clinical and prognostic information, remains an open question, with best practices that likely will vary across complex traits.

There has been mounting interest in combined phenotype statistical approaches; however, their merits for novel locus discovery and locus characterization remain largely untested in practice. Here, combined phenotype analysis of contiguous ECG traits did not identify novel loci that eluded traditional univariate analyses, despite the theoretical potential demonstrated for aSPU and related methods. Nonetheless, our evaluation of TBX5, a locus harboring multiple independent signals, suggested that combined phenotype approaches may be informative for fine-mapping. Supporting the use of combined phenotype methods to fine-mapping are methods that have been specifically developed for this challenge²¹, including fastPAINTOR. When compared with single trait fine-mapping, fastPAINTOR reduced the number of SNPs required for follow-up in order to capture 90% of the causal variants, from 23 SNPs per locus using a single trait to 12 SNPs when fine-mapping two traits simultaneously.

The lack of diversity in GWAS has long been described²², but the literature remains dominated by studies of European ancestral populations. As a result, genomics research is confined to a narrow sliver of human genetic diversity, even as the US population becomes more diverse²³. Our deliberate selection of an ancestrally diverse population enabled the identification of a novel CD36 locus, which was common only in populations of African descent. Lead SNP, rs3211938, had a large effect on QT interval, among the largest effects reported to-date¹³, although winner’s curse may be a concern¹³. Variant rs3211938, a ClinVar-indexed missense mutation known to cause CD36 deficiency, encodes a scavenger receptor central to formation and cellular uptake of long-chain fatty acids. Although CD36 and rs3211938 have been associated with a spectrum of cardiometabolic phenotypes^24–31, the most intriguing finding is the potential linkage with sudden cardiac arrest (SCA), for which QT interval prolongation increases risk³². SCA accounts for approximately 10–20% of total mortality in industrial countries³³, and several decades of research have suggested a contributory role of impaired fatty-acid uptake in cardiomyocytes¹⁴. Although genetic studies of CD36 and SCA were largely null^34,35, the use of predominantly European ancestral populations constrained evaluation of rs3211938, which is near monomorphic in all populations except those of African descent. Overall, these results highlight the potential for racially/ethnically diverse studies to provide novel biological insights beyond the reach of studies conducted in predominantly European ancestral populations.

Limitations of our study point to several promising directions for future work. First, we lacked a replication cohort, reflecting the rarity of multi-ethnic studies with high-resolution ECGs from which to derive the six contiguous ECG traits. However, this study is the largest multi-ethnic GWAS of ECGs to-date, with excellent statistical power, and we identified loci that are biologically plausible. Second, we limited our evaluation to common variants, although previous studies have demonstrated the utility of interrogating rare variants, particularly in the context of multi-ethnic studies^36,37. Our focus on common variants reflects the current limitations of combined phenotype methods for interrogating rare variants in complex samples or with summary data. Widespread interest in this approach suggests that this gap may be closed soon. Further, while this study helps address the lack of diversity in ECG trait GWAS, the small number of Chinese American participants limited our ability identify loci that were common only in populations of East Asian descent. Future efforts that further expand population racial/ethnic diversity represents an important next both for cardiac conduction studies and GWAS more broadly. Finally, in-depth fine-mapping was outside the scope of the proposed study, despite the value of multi-ethnic populations for fine-mapping³⁸. We also were unable to leverage heterogeneity in allelic effects between ethnic groups to increase statistical power, as available methods are incompatible with the multi-phenotype methods presented herein³⁹. Identification of allelic heterogeneity provides a clear impetus for future studies that leverage evidence of a shared genetic effects to further characterize the genetic architecture underlying ECG traits and, complex traits in general.

This study illustrates three strategies to improve the efficiency of locus discovery. Of these, our findings emphasize the importance of carefully constructed phenotypes and of ancestral diversity for novel locus identification. In contrast, combined phenotype methods did not enable identification of novel loci unapparent using traditional approaches, although combined phenotype studies did inform characterization of known loci. As researchers contemplate the next generation of genomics studies, increased phenotype resolution and ancestral diversity will be crucial to understanding the ever-expanding “phenome,” while ensuring equitable access to precision medicine.

Figure 1.

Illustration of the six contiguous (P wave, PR segment, QRS interval, ST segment, and TP segment) and two composite (QT interval and PR interval) ECG traits.

Nonstandard Abbreviations and Acronyms

AA

African-American

ARIC

Atherosclerosis Risk in Communities Study

aSPU

Adaptive Sum of Powered score test

CHN

Chinese-American

EA

European-American

ECG

Electrocardiogram

GWAS

Genome-wide Association Study

HCHS/SOL

Hispanic Community Health Study / Study of Latinos

HIS

Hispanic/Latino

LD

Linkage Disequilibrium

MAF

Minor Allele Frequency

MESA

Multi-Ethnic Study of Atherosclerosis

SNP

Single Nucleotide Polymorphism

PAGE

Population Architecture using Genomics and Epidemiology study