Role of Rare Variants in Systolic Blood Pressure: Analysis of Exome Chip Data in HyperGEN African Americans

Abstract

Cardiovascular disease is among the most significant health problems in the United States today, with its major risk factor, hypertension, disproportionately affecting African Americans (AAs). Although genome-wide association studies (GWAS) have identified dozens of common variants associated with blood pressure (BP) and hypertension in European Americans; these variants collectively explain less than 2.5% of BP variance, and most of the genetic variants remain yet to be identified. Here we report the results from rare variant analysis of systolic blood pressure (SBP) using 94,595 rare and low frequency variants (MAF < 5%) from the Illumina exome array genotyped in 2,045 HyperGEN AA. In addition to single-variant analysis, four gene-level association tests were used for analysis: burden and famSKAT tests using minor allele frequency (MAF) cutoffs of 1% and 5%. Gene-based methods often provided lower p-values than the single-variant approach. Some consistency was observed across these four gene-based analysis options. While neither gene-based analysis nor single variant analysis produced genome-wide significant results, the top signals, which had supporting evidence from multiple gene-based methods, were of borderline significance. Though additional molecular validations are required, six of the 16 most promising genes are biologically plausible with physiological connections to blood pressure regulation.

Introduction

Cardiovascular diseases are among the most significant health problems in the United States today, with hypertension affecting African Americans (AAs) disproportionately. Blood pressure (BP) and hypertension have been challenging phenotypes for gene discovery due to their complex nature. Although genome-wide association studies (GWAS) have successfully identified thousands of loci affecting human complex diseases, these identified loci only explain small proportions of heritability. For example, three recent GWAS consortia [1–3] identified dozens of common variants associated with blood pressure (BP) and hypertension in European Americans (EAs). However, these variants collectively explain less than 2.5% of BP variance, and most of the genetic variants remain yet to be identified [4]. The missing heritability might be partially explained by rare and low frequency variants, with minor allele frequencies (MAF) below 5% [5].

Because the power to detect individual rare variants is limited by their low frequency, the development of novel statistical approaches for the analysis of rare variants has been an active area of research. These methods are broadly classified as burden and non-burden tests. Burden tests collapse rare variants in a genomic region (typically a gene) into a single burden variable to test for the combined effects of all rare variants in the region or gene. Such tests include the cohort allelic sum test (CAST) [6], the combined multivariate and collapsing (CMC) method [7], the group-wise weighted sum test [8], and the proportion and presence tests of Morris and Zeggini [9]. Since these burden tests implicitly assume that all variants influence the trait in the same direction and with the same magnitude of effect, their power is reduced if their effects are not homogeneous [10]. Several non-burden tests have been developed that are less sensitive to heterogeneous effects [11–13]. Two one-sided statistics proposed by Ionita-Laza et al [11] quantify enrichment in risk variants and protective variants. The C-alpha test uses the distribution of genetic variation observed in cases and controls, maintaining consistent power when the target set contains both risk and protective variants [12]. The sequence kernel association test (SKAT) provides a flexible and computationally efficient approach using a score-based variance component test and applicable to both continuous and discrete traits [13]. Using the framework of linear mixed effect models, SKAT has been recently extended to handle family data, which is called the family-based SKAT (famSKAT) [14].

Here we report the results from rare variant analysis using 94,595 rare and low frequency variants from the Illumina exome array (ExomeChip) genotyped in HyperGEN African Americans (AA). We performed four gene-level association analyses of systolic blood pressure (SBP): burden and famSKAT tests with minor allele frequency (MAF) cutoffs of 1% and 5% accounting for family correlation. To account for population stratification in our AA subjects, we used principal components (PCs) of the common variants (with MAF ≥ 5%) on the ExomeChip as covariates in our analysis. A varying number of PCs (no PC, first 2 PCs and first 10 PCs) were used to evaluate the impact of population stratification on the analysis of rare variants.

Method

Study Sample

We studied African Americans (AA) in the Hypertension Genetic Epidemiology Network (HyperGEN) from the Family Blood Pressure Program [15]. The study recruited AA and EA participants at five field centers to investigate the genetic causes of hypertension and related conditions. Study participants were one of three types: 1) individuals in hypertensive sibships with at least two siblings diagnosed with hypertension; 2) random subjects, who were age-matched with hypertensive siblings; or 3) unmedicated adult offspring of one or more of the hypertensive siblings. The study obtained informed consent from participants and approval from the appropriate institutional review boards.

Genotype and Phenotype Data

A total of 2,147 HyperGEN AAs were genotyped using the Illumina exome array that covers the genome-wide exome variants. Illumina’s clustering algorithm (GenTrain version 1.0), as implemented in GenomeStudio, was used for genotype calling. To check the quality and consistency of the genotyping process, 19 blind duplicate samples were included. We removed these duplicate samples and corrected sample mix-ups and pedigree errors, resulting in 2,111 subjects. We removed monomorphic markers and SNPs with missing rate > 5% or Hardy-Weinberg p-value <10⁻⁶ and removed any genotypes with a non-Mendelian pattern of inheritance. The number of variants after quality control and exclusions was 124,759. There were 30,164 common variants (with MAF ≥ 5%) and 94,595 low frequency variants (with MAF < 5%). There were also extremely rare variants including 15,017 singletons and 9,323 doubletons.

We included 2,045 subjects with genotype, phenotype and covariates information. Systolic blood pressure (SBP) was measured (on an mmHg scale) with a DINAMAP model 1846 SX/P automated oscillometric device using a standard protocol. Briefly, after a 5-minute rest period, resting BP was measured six times, with two minutes gap between measurements. The average of 3 SBP measurements was used for each subject. In all analyses, we added 15 mmHg to their SBP values for subjects taking antihypertensive or blood pressure lowering medications. Age, sex, and body mass index (BMI) were included as covariates. In addition, a varying number of PCs derived from common variants were included as covariates, as discussed below.

Statistical Analysis

We performed single-variant analysis and the four sets of genome-wide gene-level association analysis. The four sets of gene-level association analysis are burden tests with minor allele frequency (MAF) cutoffs of 1% and 5% (Burden-1 and Burden-5) and famSKAT tests with MAF cutoffs of 1% and 5% (famSKAT-1 and famSKAT-5). We used the annotation file for the Illumina exome array data, available from the CHARGE (Cohorts for Heart and Aging Research in Genetic Epidemiology) study [16]. For single-variant analysis, we excluded singletons because false positive results may arise from an influential phenotype value of one subject. For gene-level analyses, we excluded genes with only one variant as their results were identical to single-variant analysis. We included singletons for gene-level analyses. For example, when a gene included multiple singletons, all these singletons were jointly used for our gene-based analysis.

To perform the burden tests in HyperGEN AAs, we extended the burden test by Morris and Zeggini [9] to account for family structure. For the combined effect of multiple rare variants in a genomic region, Morris and Zeggini [6] developed the burden test in a linear regression framework. For family data, a linear mixed model with a random polygenic component is commonly used to account for phenotypic correlations among related individuals. Therefore, we used a linear mixed effect modeling framework and performed our analysis with GenABEL [17] and ProbABEL [18], which uses the kinship matrix based on the pedigree structure.

Population stratification is known to create spurious associations if not properly controlled. The common practice in GWA studies has been to control for stratification by using a number of top principal components (PCs) of the GWAS markers as covariates in association analysis. However, the optimal number of PCs has not been well established even for common variant analysis, as shown by Peloso and Lunetta [19]. Therefore, to account for population stratification in our AA subjects, we used a varying number of the leading PCs as covariates in our analysis: no PC, the top 2 PCs and the top 10 PCs. Although the impact of population stratification on the analysis of rare variants has shown to be more subtle and complex [20–23], several papers have demonstrated advantages of using PCs derived from common variants (with MAF ≥ 5%) [24–27]. Therefore, we computed PCs derived from 30,164 common variants in our exome chip data. The genomic inflation factor λ was calculated as the ratio of the observed median chi-square value to the expected median chi-square value.

Results

We performed single-variant analysis and the four sets of genome-wide gene-based analysis using 79,578 low frequency variants: Burden-1, Burden-5, famSKAT-1, famSKAT-5. For a varying number of PCs, genomic inflation factors λ are shown in Table 1. Single-variant analysis exhibited no genomic inflation whether PCs were included or not (with all λ ≤ 1.002). Likewise, burden tests did not exhibit much inflation, with 2 PCs leading to the lowest inflation (λ=1.025 and 1.006 for Burden-1 and Burden-5, respectively). In contrast, SKAT analyses exhibited modest inflation, with no PCs leading to the lowest inflation (λ=1.075 and 1.094 for famSKAT-1 and famSKAT-5, respectively). Based on Table 1 and also because our subjects were of African ancestry, we decided to present the results using 2 PCs. Quantile-quantile (QQ) plots are shown in Supplemental Figure 1.

Table 1.

Genomic inflation values (λ) for the five genome-wide analysis options using each of principal components (PCs) adjustments

Number of PCs	Single-variant	Burden-1	Burden-5	famSKAT-1	famSKAT-5
0	1.002	1.028	1.007	1.075	1.094
2	0.999	1.025	1.006	1.079	1.102
10	0.999	1.041	1.010	1.107	1.119

In addition to single-variant analysis, four gene-level association tests were used to analyze the data: burden tests with a minor allele frequency (MAF) cutoff of 1% and 5% (Burden-1 and Burden-5), famSKAT tests with a frequency cutoff of 1% and 5% (famSKAT-1 and famSKAT-5). Quantile-quantile (QQ)-plots are shown in Supplemental Figure 1.

Manhattan plots for single-variant analysis and the four sets of gene-based analysis using 2 PCs are shown in Figure 1. The genome-wide Bonferroni-corrected threshold is 6.4×10⁻⁷ (=0.05/79,578) for a single-variant analysis. Gene-based tests with MAF cutoff of 1% and 5% included 11,958 and 12,895 genes, respectively; the corresponding Bonferroni-corrected thresholds were 4.2×10⁻⁶ and 3.9×10⁻⁶, respectively. Table 2 presents the top 11 low frequency variants from our single-variant analysis (with P < 6.0×10⁻⁵) and Table 3 shows the top 16 genes from the four gene-based results (with P < 2.0 ×10⁻⁴). We found no genome-wide significant results based on the Bonferroni-corrected threshold. However, we found that gene-based methods often provided lower p-values than the single-variant approach, as shown in Table 3.

Figure 1.

Manhattan plots of the single-variant analysis and the four gene-level analysis options (Burden-1, Burden-5, famSKAT-1, and famSKAT-5). For single-variant analysis, the −log₁₀(p) of each SNP was plotted at the chromosomal location of the SNP, and for each gene-level analysis, the −log₁₀(p) of each gene was plotted at the chromosomal location of the first SNP of the gene. These results were based on using 2 PCs.

Table 2.

Top low frequency variants from single-variant analysis with p-value < 6×10⁻⁵ (All these variants were exonic non-synonymous.)

Rank	SNP	Chr	BP	Gene	MAF	Beta	SE	P
1	rs138085317	4	114,290,810	ANK2	0.0042	27.3	5.6	1.2×10−6
2	rs143104022	18	30,791,983	C18orf34	0.0012	43.3	9.9	1.2×10−5
3	rs144164391	6	25,779,375	SLC17A4	0.0012	42.8	9.9	1.4×10−5
4	rs75307540	11	6,452,889	HPX	0.0093	15.9	3.8	2.9×10−5
5	rs145926130	16	1,536,575	PTX4	0.0017	35.8	8.6	3.2×10−5
6	rs35124934	2	3,744,990	ALLC	0.0342	8.1	2.0	4.3×10−5
7	rs6665210	1	145,532,823	ITGA10	0.0193	−10.7	2.6	4.4×10−5
8	rs75508009	5	111,519,758	EPB41L4A	0.0022	30.2	7.4	4.5×10−5
9	rs144714540	1	45,809,001	TOE1	0.0010	48.8	12	4.6×10−5
10	rs139543898	1	158,389,863	OR10K2	0.0005	63.4	15.6	5.0×10−5
11	rs146171474	4	122,833,191	TRPC3	0.0007	56	13.9	5.8×10−5

Genome-wide Bonferroni-corrected threshold corresponds to 6.4×10⁻⁷ (=0.05/79,578). Bold-faced genes are biologically plausible, providing physiological connection to BP regulation (as presented in Discussion); except for TRPC3, which contained only one variant, these were included in Table 3.

Table 3.

16 top genes from each gene-based analysis with p-value < 2.0 ×10⁻⁴ (The smallest P value for each gene is shown in bold.)

Group	Gene	Gene Positiona	Gene-level association p-values				Direction of single-variant associationb	Top single-variant signal (MAF < 5%)
			Burden-1	Burden-5	famSKAT-1	famSKAT-5		SNPc	MAF	P-value
1	DOPEY2	Chr 21: 37.5 Mb	8.3×10−5	2.8×10−3	8.1×10−4	2.6×10−3	+++-+++-++-(−)-++++-+++	rs142091518	0.0002	6.1×10−4
	NSUN5	Chr 7: 72.7 Mb	8.6×10−5	8.6×10−5	3.6×10−3	3.6×10−3	++++	rs77133388	0.0002	6.5×10−3
	FGF4	Chr 11: 69.6 Mb	1.2×10−4	1.2×10−4	5.6×10−4	5.6×10−4	++	rs140567674	0.0002	2.9×10−3
	OR52H1	Chr 11: 5.6 Mb	1.2×10−4	0.73	5.5×10−4	0.27	+(+)(+)+(−)(−)+++(−)(+)+	rs143228643	0.0037	4.0×10−4
	ANKRD50	Chr 4: 125.6 Mb	1.5×10−4	0.77	1.6×10−4	0.55	(−)+(+)(−)	rs140645873	0.0002	1.5×10−4
	ADC	Chr 1: 33.6 Mb	1.5×10−4	0.17	6.2×10−3	8.6×10−3	(−)++++-++	rs150169064	0.0044	0.02
	RAB3GAP1	Chr 2: 135.8 Mb	1.7×10−4	0.13	5.1×10−3	0.03	-(+)+++++(+)(−)(−)-	rs150478342	0.0017	1.7×10−4
2	FANK1	Chr 10: 127.7 Mb	0.05	1.7×10−4	0.51	9.3×10−3	---(−)--(−)--	rs141545587*	0.0005	0.01
	ANKRD6	Chr 6: 90.3 Mb	0.26	1.7×10−4	0.51	4.3×10−3	(+)(+)++	rs61736690*	0.0472	7.3×10−3
3	HPX	Chr 11: 6.5 Mb	3.2×10−4	0.41	7.1×10−5	4.0×10−3	+++++(−)(−)(−)	rs75307540	0.0093	2.9×10−5
	MARCH8	Chr 10: 46.0 Mb	0.01	0.01	1.2×10−4	1.2×10−4	--+	rs140989641	0.0022	1.0×10−4
	ARHGAP1	Chr 11: 46.7 Mb	7.5×10−3	7.5×10−3	1.4×10−4	1.4×10−4	-+-	rs11822837*	0.0044	1.5×10−4
	CD3E	Chr 11: 118.2 Mb	4.6×10−4	4.6×10−4	1.5×10−4	1.5×10−4	+++	rs140639753	0.0015	1.1×10−4
4	MOCS1	Chr 6: 39.9 Mb	0.11	2.3×10−5	0.01	1.4×10−5	++-(+)+-	rs7762875	0.0208	6.2×10−5
	ALLC	Chr 2: 3.7 Mb	0.95	1.5×10−4	0.03	1.7×10−5	-+--+(+)(+)-(+)	rs35124934	0.0342	4.3×10−5
	ITGA10	Chr 1: 145.5 Mb	0.15	8.8×10−3	1.3×10−3	8.6×10−5	-+(−)+-+-(+)--++(−)++-(−)-	rs6665210	0.0193	4.4×10−5

Four gene-level association tests were used to analyze the data: burden tests with minor allele frequency (MAF) cutoffs of 1% and 5% (Burden-1 and Burden-5), famSKAT tests with MAF cutoffs of 1% and 5% (famSKAT-1 and famSKAT-5). Bold-faced genes are biologically plausible, providing physiological connection to blood pressure regulation (as presented in Discussion).

^aGene position is defined according to hg19, GRCh37 (Genome Reference Consortium Human Reference 37).

^bDirection of single-variant association statistics for variants with MAF< 5%. Variants in parentheses have MAF >1%. Bonferroni-corrected threshold corresponds to 4.2×10⁻⁶ and 3.9×10⁻⁶ for MAF <1% and MAF <5%, respectively.

^cAll these most strongly associated variants were exonic non-synonymous.

The three variants with * were also splicing.

We found modest consistency across these four gene-based analysis options. The results for the burden test and famSKAT test within each MAF threshold were moderately correlated (Spearman correlation = 0.55). The MAF thresholds (1% and 5%) made more difference for the burden tests than famSKAT tests (Spearman correlation = 0.50 and 0.72, respectively). The scatterplots of pairwise −log₁₀(p) values across four gene-based analysis options are presented in Supplemental Figure 2.

We found that all 16 top genes in Table 3 indeed had supporting evidence from multiple gene-based methods. To further evaluate the relative performance of each gene-based method, we classified these genes in Table 3 into 4 groups. Note that the low frequency variants (with 1% ≤ MAF < 5%) are presented in parenthesis in the “direction of single-variant association” column. For genes in group 1, the Burden-1 approach provided the strongest evidence; at these genes, rare variants had mostly unidirectional effects. We found that famSKAT also provided corroborating evidence for some of those genes (notably NSUN5 and FGF4), which is assuring. When the low frequency variants had the opposite effects as compared to rare variants (as OR52H1), the Burden-5 approach provided much weaker evidence than Burden-1. In contrast, when the low frequency variants had the same effects as rare variants, the Burden-5 approach provided stronger evidence than Burden-1, as shown in group 2. We found that famSKAT indeed performed better on genes with multidirectional variants, as shown in groups 3 and 4 (notably on HPX).

Some genes in Table 3 had the most strongly associated single variant with the same magnitude of p-value as the gene-level results. To assess if the signal from gene-level analyses is primarily driven by the one single variant or aggregated evidence from multiple low-frequency variants, we performed conditional analysis by excluding this most strongly associated variant and using it as an additional covariate. Supplemental Table 1 provides p-values obtained by our conditional analysis. The signal for DOPEY2 appears to reflect aggregate evidence from multiple low-frequency variants. Five other genes including NSUN5 also have a hint of aggregated effects, whereas the remaining genes have no significant conditional p-values, indicating their signals mostly driven by one variant.

To improve the selection of variants to be included in the gene-level tests, we also considered functional annotation information. It turned out that most (90,875 out of 94,595) low frequency variants were non-synonymous, stop-gain/loss, splice, or missense variants. Therefore, when we repeated all four gene-level analyses with these functional variants only, the results were very similar (Spearman correlation = 0.97) as shown in the first row of Supplemental Figure 3. Roughly half (N=41,393) of the low frequency variants were stop-gain/loss, splice or missense and predicted to be damaging by 2 of the 4 algorithms in dbNSFP (Mutation Taster, Polyphen HDIV, SIFT or LRT). We further performed four gene-level analyses with these damaging variants and found a subset of genes with lower p-values (shown in the second row of Supplemental Figure 3). The most notable one was MOCS1, where Burden-5 result was enhanced (P = 4.5 ×10⁻⁶). Our results illustrate that excluding neutral variants may help improve the power of gene-level tests.

Although our focus was on low frequency variants, we also performed single-variant analysis on common variants. Among 30,164 common variants, our strongest signal was observed at rs11711441 on chromosome 3 (with P = 1.7 ×10⁻⁵). Out of 60 known BP variants identified by the 9 articles [3, 28–35], only 13 were available in our ExomeChip data. Among these 13 known variants, the smallest p-value from our analysis was observed at rs3184504 in SB2B3 (with P = 0.0029); the effect was the same direction as [3], but with lower MAF (0.057). Our results may not be a surprise given a limited number of common variants in the Exome Chip data.

Discussion

We performed analysis of systolic blood pressure (SBP) using 79,578 rare and low frequency variants from the Illumina exome array genotyped in 2,045 HyperGEN AAs. While neither gene-based analysis nor single variant analysis produced genome-wide significant results, the most promising genes from each of the 4 methods had supporting evidence from some of the other gene-based methods. Of the genes for which the burden approach provided strongest evidence, three genes are of particular interest as potential BP candidate genes: NSUN5 (NOP2/Sun domain family), ADC (arginine decarboxylase), and RAB3GAP1 (RAB3 GTPase activating protein subunit 1). All 3 genes had corroborating evidence from SKAT, as well. As the burden approach provided more significant evidence, the SNP effects in these genes are likely homogeneous. NSUN5 is shown to be deleted in Williams-Beuren syndrome [36], a multi-system developmental disorder whose symptoms include hypertension [37]. ADC contains a conformational change-inducing missense mutation (substituting a polar amino acid, serine, for a non-polar one, alanine) that has been previously associated with SBP [38]. RAB3GAP1 has been previously identified to be associated with sudden cardiac death in the MetaboChip study [39].

Of the genes with strong evidence from famSKAT analysis, three have potential connections to BP: HPX (Hemopexin), ARHGAP1 (Rho GTPase activating protein 1), and ALLC (allantoicase). HPX is a plasma glycoprotein that transports heme to the liver, and it is a major BP candidate gene owing to its role in regulating expression of the angiotensin II type I receptor AGTR1 and modulating sensitivity to angiotensin [40] as well as its identification as a risk factor for late renal graft failure [41]. ARHGAP1 is a guanosine-triphosphate metabolizing enzyme that plays a role in the regulation of tubule formation, an important early step in angiogenesis. ALLC is involved in the uric acid degradation pathway; uric acid levels have been associated with BP [42].

Single-variant analysis provided evidence for two additional potential candidate genes: ANK2 (ankyrin 2) and TRPC3 (transient receptor potential cation channel). ANK2 is a member of the ankyrin family of proteins and is required for the targeting and stability of the Na⁺/Ca²⁺ exchanger in cardiomyocytes. Mutations in this gene lead to long-QT syndrome and cardiac arrhythmia [43], and ANK2 was located in a 1-lod support interval linked to resting heart rate [44]. TRPC3 is a renal Ca²⁺ channel expressed in principal cells of the collecting duct. Expression levels of TRPC3 have been associated with systolic hypertension in humans [45] and are higher in spontaneously hypertensive rats than in normotensive Wistar Kyoto rats [46].

To evaluate the impact of population stratification on the analysis of rare variants, we used a varying number of leading principal components (PCs) as covariates in our analysis. We found that type I error rates were well controlled for the single-variant and burden approaches, whereas the famSKAT approach exhibited modest inflation, as also noted by the famSKAT developers [14]. Furthermore, a recent article by Zawistowski et al [23] also reports higher levels of inflation from the other non-burden tests including the C-alpha test [12] and the original SKAT [13]. These non-burden tests have greater power to detect association in genes containing rare variants with heterogeneous effects but may be more vulnerable to population stratification. Given the lack of significant findings due to low power to detect rare variants, consistent findings from multiple gene-based methods including both burden and non-burden approaches with different thresholds is reassuring. These findings are compelling candidates for external validation.