Identifying Blood Pressure Loci Whose Effects Are Modulated by Multiple Lifestyle Exposures

Abstract

Although multiple lifestyle exposures simultaneously impact blood pressure (BP) and cardiovascular health, most analysis so far has considered each single lifestyle exposure (e.g., smoking) at a time. Here, we exploit gene-multiple lifestyle exposure interactions to find novel BP loci. For each of 6,254 Framingham Heart Study participants, we computed lifestyle risk score (LRS) value by aggregating the risk of four lifestyle exposures (smoking, alcohol, education, and physical activity) on BP. Using the LRS, we performed genome-wide gene-environment interaction analysis in systolic and diastolic BP using the joint 2 degree of freedom (DF) and 1 DF interaction tests. We identified one genome-wide significant (P-value < 5 × 10⁻⁸) and 11 suggestive (P-value < 1 × 10⁻⁶) loci. Gene-environment analysis using single lifestyle exposures identified only one of the 12 loci. Nine of the 12 BP loci detected were novel. Loci detected by the LRS were located within or nearby genes with biologically plausible roles in the pathophysiology of hypertension, including KALRN, VIPR2, SNX1, and DAPK2. Our results suggest that simultaneous consideration of multiple lifestyle exposures in gene-environment interaction analysis can identify additional loci missed by single lifestyle approaches.

Introduction

Blood pressure (BP) is a heritable (Ehret, 2010) but modifiable risk factor for cardiovascular diseases (CVD) with major global health and economic burden (Forouzanfar et al., 2017). Characterizing the genetic architecture of BP is critical to advancing our understanding of underlying biological mechanisms, enhancing risk prediction, and developing targeted lifestyle interventions and drugs (Timpson, Greenwood, Soranzo, Lawson, & Richards, 2018). Genome-wide association studies (GWAS) have identified at least 901 loci associated with BP, which are estimated to account for 11.2% of the genetic variance in BP (Evangelou et al., 2017); a large proportion of heritability in BP remains unexplained. This leaves room for enhanced locus discovery by exploring alternative approaches.

One such approach is to consider gene-lifestyle interactions in BP, given that several lifestyle exposures (e.g., smoking status) have been shown to modulate genetic effects on complex traits such as BP (Parnell et al., 2014; Rao et al., 2017). By accounting for gene-lifestyle interactions in BP, one can find genetic loci whose effects vary by lifestyle exposure. Such genetic loci may otherwise be undetected in traditional genetic main effect GWAS (Rao et al., 2017). For example, recent genome-wide interaction analyses performed by the CHARGE Gene-Lifestyle Interactions Working Group identified several novel loci with biologic plausibility for their involvement in BP and lipid homeostasis (Bentley et al., 2019; de Vries et al., 2019; Feitosa et al., 2018; Kilpeläinen et al., 2019; Sung et al., 2019, 2018). Finding such loci is of value for gaining insight to the complex biological underpinnings of BP (Laville et al., 2019).

To date, most genome-wide gene-lifestyle interaction analyses have primarily considered gene interactions with a single lifestyle exposure, such as alcohol (de Vries et al., 2019; Feitosa et al., 2018; Simino, Sung, Kume, Schwander, & Rao, 2013), smoking (Bentley et al., 2019; Sung, de las Fuentes, Schwander, Simino, & Rao, 2015; Sung et al., 2018), and education (Basson et al., 2014). Yet, the genetic effects of some loci may be simultaneously influenced by multiple lifestyle exposures given that cardiovascular health (including BP) is concurrently impacted by multiple lifestyle exposures (Folsom et al., 2011; van Dam, Li, Spiegelman, Franco, & Hu, 2008). In fact, the current strategy for improving cardiovascular health in the general population promotes multiple healthy lifestyle behaviors simultaneously (Lloyd-Jones et al., 2010). We may be able to further enhance discovery of novel BP loci whose effects are modulated by multiple lifestyle exposures by using genome-wide approaches that account for interactions with more than one lifestyle exposure.

To identify loci that interact with multiple environmental factors, a Bayesian approach has been developed, but its implementation is burdened by high computational requirements and an inability to account for family relatedness (Moore et al., 2018). An alternative approach would be to aggregate multiple lifestyle exposures into a composite score that captures overall risk for unfavorable cardiovascular outcomes due to unhealthy lifestyles. This approach has been applied in epidemiological studies (Nayor, Enserro, Vasan, & Xanthakis, 2016; Ogunmoroti et al., 2017), used to examine how genetic risk for CVD and BP is modified by a composite score derived from multiple lifestyle exposures (Pazoki et al., 2018), and recently compared to other approaches for jointly testing multiple interactions involving several candidate genetic variants and exposures (Kim et al., 2019). However, it is yet to be used in genome-wide gene-lifestyle interaction analysis for finding BP loci.

In this study, we report on our findings using a risk score aggregating multiple lifestyle exposures, hereafter ‘Lifestyle risk score’ (LRS), when identifying BP loci. We selected four lifestyle exposures with well-documented effects on BP and cardiovascular health (Chobanian et al., 2003): smoking status, alcohol intake, level of educational attainment (a proxy for socio-economic status), and physical activity. We performed genome-wide interaction studies on BP using both quantitative and discretized LRS. Additionally, we examined whether our analysis using the LRS can identify additional loci not detected through analysis using each of single lifestyle exposures separately.

Methods

Study Population

In this study, we used the Framingham Heart Study (FHS) data, obtained from the database of genotypes and phenotypes (dbGaP). FHS study participants are made up of three cohorts; the Original Cohort recruited in 1948, the Offspring Cohort recruited in 1971, and the Third Generation Cohort recruited in 2002. Across these three cohorts, we selected contemporaneous visits, namely, the 26^th visit for the Original Cohort, the 7^th visit for the Offspring Cohort, and the 1^st visit for the Third Generation Cohort.

In this study, we considered FHS participants of European ancestry that were between 18 and 80 years (to avoid growth or age related changes to BP) (Shankar, Eckert, Saha, Tu, & Pratt, 2005; Wei, 1992), and had non-missing genotype data and phenotype data, including age, sex, SBP and/or DBP, anti-hypertensive medication status, and the four component lifestyle exposures (smoking status, alcohol intake, education, and physical activity). After applying these study inclusion criteria, there were 6,254 and 6,253 participants for SBP and DBP, respectively, and participants belonged to either the Offspring or Third Generation Cohort.

Genotype data

We removed genotyped single nucleotide polymorphisms (SNPs) from FHS data with Hardy-Weinberg equilibrium P-values < 10⁻⁶ or call rates < 90%, and imputed SNPs with imputation quality measures < 0.3. We further removed SNPs with minor allele frequency ≤ 1% or, if the product of minor allele count and imputation quality was ≤ 20. Appendix Table S1 shows a summary of the resulting analysis SNP sets, composed of ∼2.4 million SNPs.

Phenotype data

SBP and DBP were measured with a consistent protocol. For each individual, BP values represent an average of three measurements, one taken by a nurse/technician and two taken by a physician. For individuals on anti-hypertensive medications, BP values were adjusted by adding 15 mmHg to SBP and 10 mmHg to DBP (Tobin, Sheehan, Scurrah, & Burton, 2005).

Lifestyle exposures and LRS

We selected the following four lifestyle exposures: smoking status, alcohol intake, level of educational attainment (a proxy for socio-economic status), and physical activity.

We considered two dichotomous variables of each of the first three lifestyle exposures and one variable for physical activity. Table 1 shows the derivation and coding for each of the dichotomous variables representing the four lifestyle exposures. For example, for smoking status, the variable “Smoke now” (Smk_now), reflects whether the individual was a smoker at the time of the clinic visit, while “Smoke ever” (Smk_ever), reflects whether an individual has ever smoked (in the past or at the time of the clinic visit).

Table 1.

Lifestyle exposure variables used in this study

Lifestyle exposure	Variable		Original variable(s) in FHS	Modification/Remark	Values*
Smoking	1.	Smk_now	Smoke regularly in the past year?	0: former and never smokers 1: current smokers only	Smk_now (0) Smk_now (1)
	2.	Smk_ever	Ever smoke cigarettes regularly?	0: never smokers only 1: former and current smokers	Smk_ever (0) Smk_ever (1)
Alcohol	3.	Alc_heavy	Total number of alcoholic drinks/week	0: consumed 0 or 1 – 7 drinks/week	Alc_heavy (0)
				1: consumed > 7 drinks/week	Alc_heavy (1)
	4.	Alc_moderate	Total number of alcoholic drinks/week	0: consumed 0 or > 7 drinks/week	Alc_moderate (0)
				1: consumed 1 – 7 drinks/week	Alc_moderate (1)
Education	5.	Edu_somecol	Post-secondary education? (Education level in categories/years)	0: if no post-secondary education (i.e., no schooling, grades 1 – 11, high school grad or GED or years of education ≤ 12).	Edu_somecol (0)
				1: if there is some form of post-secondary education (i.e., some college, technical school, associate degree, bachelors degree, or graduate/professional degree or years of education ≥ 13)	Edu_somecol (1)
	6.	Edu_gradcol	Graduate college? (Education level in categories/years)	0: if no graduate college degree (i.e., all the levels below bachelors degree defined above).	Edu_gradcol (0)
				1: if bachelors/graduate/professional degree or ≥ 16 years of education.	Edu_gradcol (1)
Physical activity	7.	PA_active	Hours spent in moderate activity, and hours spent in heavy activity	Sum of time spent in moderate activity and twice the time spent in heavy activity, then split by lower quartile (Q1=3); 0: ≥ 3 hours of time equivalent in moderate and heavy physical activity 1: < 3 hours of time equivalent in moderate and heavy physical activity.	PA_active (0) PA_active (1)

^*Value assignment across variables was done in one of two ways; for Smk_now, Smk_ever, Alc_heavy and PA_active, 0 was assigned to the unexposed group and 1 to the exposed group, while for Alc_moderate, Edu_somecol and Edu_gradcol, the unexposed group was assigned a value of 1 and the exposed group a value of 0.

For analyses involving a composite of multiple lifestyle exposures, the LRS was computed in two main steps:

First, risk was assigned to each value of a lifestyle exposure based on its well-established effect on BP or cardiovascular health (Chobanian et al., 2003; Loucks, Abrahamowicz, Xiao, & Lynch, 2011; WHO, 2013). As presented in Table 2, three risk scores were used, no risk (0), low risk (1), and high risk (2). High risk scores are associated with unfavorable cardiovascular health outcomes. For alcohol intake, risk assignment reflects reported J-shaped associations indicating the cardio-protective effect of moderate alcohol consumption compared to abstinence, based on extensive data including healthy individuals and CVD patients (Costanzo et al., 2019; O’Keefe, Bybee, & Lavie, 2007). Unlike the other lifestyle exposures, physical activity has only two risk levels (no risk and low risk) because the effect of any level of activity on BP (as compared to sedentariness) is much more pronounced than amount of activity among active individuals (Diaz & Shimbo, 2013). Moreover, there is no consensus on a cutoff to distinguish between active and very active individuals (with discernible differences in the effects on BP); this problem is compounded by the known overestimation (or underestimation) of self-reported physical activity values (Prince et al., 2008).

Table 2.

Component lifestyle exposures and risk assignment for computing Lifestyle risk score (LRS)

Component variable	No risk (0)	Low risk (1)	High risk (2)
Smoking	Never	Former	Current
Alcohol intake (drinks/week)	Modest (1 – 7)	Abstinence (0)	Heavy (> 7)
Education	College degree	Some college	None
Physical activity	Active	Inactive

Subsequently, the quantitative LRS (QLRS) was computed by summing across risk scores of all four lifestyle exposures. We also dichotomized the LRS, as is common practice to harmonize heterogeneous quantitative exposure variables in multi-cohort collaborations (Palla, Higgins, Wareham, & Sharp, 2010). The dichotomous LRS (DLRS), was derived by splitting QLRS by the median threshold of 2; individuals with QLRS values < 2 were assigned to the unexposed (favorable lifestyle) group, while individuals with QLRS values ≥ 2 were assigned to the exposed (unfavorable lifestyle) group. The favorable lifestyle group includes individuals who were at no risk for any of the four exposures or at low risk for only one of the exposures. Individuals with all other combinations of risk levels for the four exposures would fall into the unfavorable lifestyle group.

Statistical analysis

Genome-wide gene-LRS interaction scan: we fit a linear mixed effect model that jointly models the genetic, lifestyle, and interaction effects (Kraft, Yen, Stram, Morrison, & Gauderman, 2007), while adjusting for covariates and potential confounders (e.g., lifestyle-covariate interactions) (Keller, 2014). The model is

$$y={\beta }_0+{\beta }_{E}E+{\beta }_{G}G+{\beta }_{GE}G*E+{\beta }_{cov}{X}_{cov}+\epsilon$$ (1)

where y is the BP trait (SBP or DBP), E is the lifestyle exposure (QLRS or DLRS), G is the dosage of the genetic variant, and X_cov is the vector of covariates including age, sex, lifestyle exposure interactions with age (age*E) and sex (sex*E). ${\beta }_0$ is the intercept, ${\beta }_E$ is the environmental main effect, ${\beta }_G$ is the genetic main effect, ${\beta }_{GE}$ is the gene-environment interaction effect, ${\beta }_{cov}$ are the covariate effects, and $\epsilon$ is the error term. From the model output, we calculated two test statistics which follow a χ² distribution under the null; a 2 degree of freedom (DF) joint test of ${\beta }_G$ and ${\beta }_{GE}$ (H₀: ${\beta }_G={\beta }_{GE}=0$) and a 1 DF test of ${\beta }_{GE}$ (H₀: ${\beta }_{GE}=0$).

Genome-wide genetic effect scans: to further assess the extent to which signals detected using Model (1) above would be missed in the absence of the interaction term, we performed two additional genome-wide scans using main effects models that adjust for the effect of exposures but do not include a gene-environment interaction term, namely, Models (2) and (3) as defined below:

$$y={\beta }_0+{\beta }_{E}E+{\beta }_{G}G+{\beta }_{cov}{X}_{cov}+\epsilon$$ (2)

$$\begin{gathered} y={\beta }_0+{\beta }_{E1}smokingstatus+{\beta }_{E2}alcoholintake+{\beta }_{E3}education+ \\ {\beta }_{E4}physicalactivity+{\beta }_{G}G+{\beta }_{cov}{X}_{cov}+\epsilon \end{gathered}$$ (3)

where model terms are as defined for Model (1), with the exception that there is no gene-environment interaction term in Models (2) and (3). Moreover, in Model (3), in lieu of the aggregate LRS, each exposure variable (i.e., smoking status, alcohol intake, education, and physical activity) is separately adjusted (i.e., ${\beta }_{E1},{\beta }_{E2},{\beta }_{E3},{\beta }_{E4}$). Also, in Model (3), covariates adjusted for include age, sex, as well as age interactions with each lifestyle exposure (i.e., age*smoking status, age*alcohol intake, age*education, and age*physical activity) and sex interactions with each lifestyle exposure (i.e., sex*smoking status, sex*alcohol intake, sex*education, and sex*physical activity). Here, based on the output from the models, we calculated a 1 DF test of the genetic main effect ${\beta }_G$ (H₀: ${\beta }_G=0$).

Genome-wide gene-single lifestyle exposures scan: we performed genome-wide scans fitting Model (1) as defined above, using each of seven single lifestyle exposure variables (defined in Table 1) as E in Model (1). This genome-wide scan was performed to determine whether the LRS identifies additional loci not detected through analysis using each of single lifestyle exposures separately. As described above in the genome-wide gene-LRS scan, we calculated a 2 DF joint test and 1 DF interaction test.

We accounted for family relatedness in the FHS dataset by applying the genome-wide rapid association using mixed model and regression (GRAMMAR) approach (Aulchenko, de Koning, & Haley, 2007). This approach involves obtaining pedigree-adjusted residuals by fitting a polygenic model on the phenotype (BP traits) using the kinship matrix as a random component. The GRAMMAR approach was applied using the GenABEL package in R (Aulchenko et al., 2007). The resulting pedigree-adjusted residuals were used as the phenotype in the models described above. We obtained effect estimates using the ProbABEL package from the GenABEL suite of programs (Aulchenko, Struchalin, & van Duijn, 2010).

In all, we performed 18 genome-wide analyses using Model (1) to determine whether using an aggregate of multiple lifestyle exposures can identify additional BP loci relative to using individual lifestyle exposures: 2 phenotypes (SBP and DBP) x 9 lifestyle exposure variables (7 individual components of the LRS + QLRS + DLRS). Further, we performed six more genome-wide scans, using Models (2) and (3) to assess the extent to which the interaction term contributed to discovering signals.

We identified SNPs as significant using the threshold of P-value < 5 × 10⁻⁸, and as suggestive using the threshold of P-value < 1 × 10^-6. For each suggestive/significant association, a locus was defined as a cluster of SNPs within 500 kb of the index SNP (i.e., the SNP with the lowest P-value in the region). We considered loci as novel if component SNPs were not within 500 kb of previously known BP loci in published literature between 2011 and 2019 (reference list in Supporting information S1), and were not detected in the CHARGE Gene-Lifestyle Interactions Working Group’s large-scale studies on gene-smoking (Sung et al., 2018) and gene-alcohol (Feitosa et al., 2018) interactions for BP. Further, we annotated suggestive or significant SNPs using FUMA GWAS (Functional Mapping and Annotation of Genome-Wide Association Studies) (Watanabe, Taskesen, Bochoven, & Posthuma, 2017). Annotation was focused on RegulomeDB categorical scores (Boyle et al., 2012), which indicate evidence for the presence of regulatory elements (e.g., enhancers, promoters, insulators or transcription binding factors), and CADD (Combined Annotation Dependent Depletion) scores (Kircher et al., 2014), which predict whether the functional consequence of a variant is likely to be deleterious. Lastly, using the GWAS atlas tool (https://atlas.ctglab.nl/; Watanebe et al. 2019), we explored whether suggestive or significant SNPs detected by the gene-LRS genome-wide scan have been previously reported in the literature as associated with a BP or other cardiovascular trait at nominal significance (i.e., P-value < 0.05 for the marginal genetic effect).

Results

Descriptive summary statistics for the study sample are presented in Table 3. For QLRS that ranged from 0 – 7, most individuals had intermediate values of the score and with fewer individuals in the tails (Figure 1). Less than a third of the 6,254 eligible participants in the study sample fell into the favorable lifestyle (unexposed) group of the DLRS and were on average younger, more likely female, and had lower BP than those in the exposed group.

Table 3.

Descriptive statistics by lifestyle risk score in FHS study sample

	QLRS	DLRS
		Unexposed	Exposed
Sample size	6254	1924	4330
Age (years)	48.4±13.6	44.1±12.5	50.3 ±13.6
% Male	46.9	43.0	48.6
% Taking anti-hypertensive medications	18.1	10.9	21.3
SBP (mmHg)	122.8±18.9	117.7±16.7	125.0±19.4
DBP (mmHg)	76.7±10.4	75.3±10.0	77.3±10.5

Sample size shown is for individuals with SBP phenotype data and is the basis for other descriptive statistics shown. Descriptive statistics are very similar for DBP phenotype data with sample size differing by a single individual (N=6,253).

Figure 1.

Distribution of lifestyle risk score (LRS) in the FHS data. The quantitative LRS (QLRS), created by summing risks across the four lifestyle exposures, ranged between 0 and 7. The dichotomous LRS (DLRS), created by splitting QLRS by the median of its distribution, was composed of two groups: favorable lifestyle group (QLRS < 2) and unfavorable lifestyle group (QLRS ≥ 2).

QQ plots are presented in appendix Figures S1 and S2. For SBP, genomic control values (λ) ranged from 0.962 – 1.027 for the joint 2 DF test and from 0.952 – 1.036 for the interaction 1 DF test. For DBP, λ ranged from 0.882 – 0.923 for the joint 2 DF test and from 0.966 – 1.022 for the interaction 1 DF test. There was no indication of genomic inflation in any of the 18 analysis sets.

Figure 2 shows a Venn diagram summarizing the numbers of detected loci that were unique to the QLRS, the DLRS, and the single lifestyle exposures. While single lifestyle exposures detected substantially more loci (26 loci in total compared to seven by the QLRS and five by the DLRS), the QLRS and DLRS detected non-overlapping loci that would have been missed in the absence of an aggregate LRS analysis. Association results from the LRS analysis are summarized in Table 4 (Manhattan plots in appendix Figure S3). Also, for associations detected using the LRS analysis, Table 5 presents P-values for the genetic effect when the marginal effect of exposures are adjusted for but interaction terms are absent (i.e., based on Models (2) and (3)). The LRS analysis identified three genome-wide significant and 18 suggestive SNPs, representing 12 loci in all. Eleven of these 12 loci detected by the QLRS and DLRS analyses were not captured by analyses using single lifestyle exposures. With the exception of the one locus that was found in common by QLRS and Smk_ever analysis for SBP, there was generally no overlap in loci found by analysis using each LRS and the seven single lifestyle exposures.

Figure 2.

Venn diagram showing the number of loci detected, either uniquely or in common, by analysis using the QLRS, DLRS, and seven single lifestyle exposures.

Table 4.

Twenty-one SNPS that showed suggestive/significant associations with BP in gene-LRS analysis in FHS

Locus	Chr:position†	SNP	Genomic location¶	Effect Allele	EAF	BP trait	Exposure	${\mathit{\beta }}_{\mathit{G}}$	${\mathit{S}\mathit{E}}_{\mathit{G}}$	${\mathit{P}}_{\mathit{G}}$	${\mathit{\beta }}_{\mathit{G}\mathit{E}}$	${\mathit{S}\mathit{E}}_{\mathit{G}\mathit{E}}$	${\mathit{P}}_{\mathit{G}\mathit{E}}$	${\mathit{P}}_{\mathit{J}\mathit{o}\mathit{i}\mathit{n}\mathit{t}2\mathit{d}\mathit{f}}$
1	2:134312960	rs1368091	Intronic NCKAP5	G	0.281	DBP	DLRS	−1.15	0.28	2.85E-05	1.73	0.35	7.23E-07	4.16E-06
2	3:123923233	rs7652065	Intronic KALRN	A	0.376	SBP	DLRS	−1.67	0.38	1.48E-05	2.43	0.50	8.82E-07	4.03E-06
	3:123923922	rs1444768	Intronic KALRN	G	0.376	SBP	DLRS	−1.66	0.38	1.48E-05	2.43	0.49	8.74E-07	4.01E-06
	3:123927014	rs880000	Intronic KALRN	G	0.377	SBP	DLRS	−1.62	0.38	2.00E-05	2.40	0.49	8.63E-07	4.32E-06
	3:123930542	rs1444757	Intronic KALRN	C	0.377	SBP	DLRS	−1.62	0.38	1.99E-05	2.40	0.49	9.13E-07	4.54E-06
3	3:143028187	rs17636599	Intronic SLC9A9	T	0.011	SBP	QLRS	3.39	1.61	3.59E-02	−2.41	0.51	2.28E-06	1.80E-07
4	4:180139179	rs10520420	Intergenic	T	0.988	DBP	DLRS	3.98	0.74	9.23E-08	−4.06	1.03	7.71E-05	6.34E-07
	4:180151019	rs4146838	Intergenic	G	0.988	DBP	DLRS	4.07	0.73	3.01E-08	−4.16	1.02	4.63E-05	2.13E-07
	4:180164081	rs969219	Intergenic	A	0.988	DBP	DLRS	4.19	0.72	6.32E-09	−4.29	1.01	2.31E-05	4.69E-08
	4:180178378	rs11131920	Intergenic	G	0.010	DBP	DLRS	−5.10	0.67	4.01E-14	5.29	0.99	8.95E-08	3.71E-13
	4:180179576	rs11131921	Intergenic	T	0.010	DBP	DLRS	−5.09	0.67	4.19E-14	5.29	0.99	9.14E-08	3.90E-13
5	7:158848821	rs3793217	Intronic VIPR2	A	0.865	DBP	QLRS	−1.14	0.33	6.55E-04	0.62	0.12	1.44E-07	3.34E-07
6	8:25374910‡	rs2012485	Intergenic CDCA2	C	0.961	SBP	QLRS	−4.31	1.02	2.62E-05	1.80	0.37	8.23E-07	4.51E-06
7	8:54274100	rs7001769	Intergenic OPRK1	T	0.828	SBP	QLRS	2.23	0.51	1.18E-05	−0.35	0.19	6.20E-02	7.63E-07
	8:54277247	rs7465458	Intergenic OPRK1	A	0.828	SBP	QLRS	2.24	0.51	1.03E-05	−0.35	0.19	5.81E-02	6.94E-07
8	11:11395489§	rs10047474	Intronic GALNT18	T	0.457	DBP	QLRS	1.81	0.34	0.00E+00	−0.48	0.12	5.80E-05	4.47E-07
9	11:49202187‡	rs7124497	Intronic FOLH1	G	0.950	SBP	DLRS	−1.06	0.97	2.73E-01	4.41	1.16	1.47E-04	7.40E-07
10	12:133441378	rs4303268	Intronic CHFR	G	0.892	SBP	QLRS	−3.68	0.83	9.67E-06	1.43	0.29	7.73E-07	4.09E-06
11	15:64315336	rs12442060	Intronic DAPK2	C	0.635	SBP	DLRS	1.68	0.40	2.53E-05	−2.55	0.51	5.39E-07	3.15E-06
	15:64316458	rs172995	Intronic DAPK2	A	0.715	SBP	DLRS	1.80	0.44	3.35E-05	−2.76	0.55	6.52E-07	3.93E-06
12	18:30053150	rs11081767	Intergenic GAREM	A	0.466	DBP	QLRS	1.11	0.24	2.00E-06	−0.43	0.09	7.14E-07	2.24E-06

^†Chr:position: based on human genome build 37.

^‡Within boundaries of known loci previously reported to be associated with BP (in published literature between 2011 and 2019; see S?).

^§Within boundaries of locus detected in CHARGE Gene-Lifestyle Interactions Working Group’s large-scale study on gene-alcohol interactions on BP (Feitosa et al., 2018).

^¶Genomic location: sourced from FUMA GWAS (https://fuma.ctglab.nl). Nearest mapped gene is indicated for intergenic SNPs, while gene located within is indicated for intronic SNPs. No gene is specified for 5 SNPs because only unmapped/uncharacterized genes are located within 500 kb of these SNPs.

Genome-wide significant SNPs at P-value < 5 × 10⁻⁸ are in bold font. All other SNPs are suggestive at P-value < 1 × 10⁻⁶.

Abbreviations: Chr, chromosome; QLRS, quantitative lifestyle risk score; DLRS, dichotomous lifestyle risk score; EAF, effect allele frequency; ${\mathit{\beta }}_{\mathit{G}}$, genetic main effect; ${\mathit{S}\mathit{E}}_{\mathit{G}}$, standard error for genetic main effect; ${\mathit{P}}_{\mathit{G}}$, P-value for genetic main effect; ${\mathit{\beta }}_{\mathit{G}\mathit{E}}$, gene-environment interaction effect; ${\mathit{S}\mathit{E}}_{\mathit{G}\mathit{E}}$, standard error for gene-environment interaction effect; ${\mathit{P}}_{\mathit{G}\mathit{E}}$, P-value of 1 degree of freedom (DF) interaction test; ${\mathit{P}}_{\mathit{J}\mathit{o}\mathit{i}\mathit{n}\mathit{t}2\mathit{d}\mathit{f}}$, P-value of 2 DF Joint test.

Table 5.

P-values for the genetic effect using refined main effects models, which omits the interaction term but adjusts for the marginal effect of the LRS (Model (2)) or the marginal effects of the four component lifestyle variables (Model (3)).

Chr:position†	SNP	BP trait	Exposure	${\mathit{P}}_{\mathit{G}}$Model 2	${\mathit{P}}_{\mathit{G}}$Model 3
2:134312960	rs1368091	DBP	DLRS	7.59E-01	8.37E-01
3:123923233	rs7652065	SBP	DLRS	9.56E-01	9.13E-01
3:123923922	rs1444768	SBP	DLRS	9.54E-01	9.11E-01
3:123927014	rs880000	SBP	DLRS	8.75E-01	8.34E-01
3:123930542	rs1444757	SBP	DLRS	8.87E-01	8.49E-01
3:143028187	rs17636599	SBP	QLRS	1.20E-02	1.42E-02
4:180139179	rs10520420	DBP	DLRS	9.21E-02	8.20E-02
4:180151019	rs4146838	DBP	DLRS	8.77E-02	7.75E-02
4:180164081	rs969219	DBP	DLRS	8.20E-02	7.21E-02
4:180178378	rs11131920	DBP	DLRS	5.01E-02	4.25E-02
4:180179576	rs11131921	DBP	DLRS	5.03E-02	4.27E-02
7:158848821	rs3793217	DBP	QLRS	5.71E-02	7.41E-02
8:25374910	rs2012485	SBP	QLRS	8.19E-01	7.91E-01
8:54274100	rs7001769	SBP	QLRS	3.36E-06	5.78E-06
8:54277247	rs7465458	SBP	QLRS	3.36E-06	5.61E-06
11:11395489	rs10047474	DBP	QLRS	1.06E-03	1.64E-03
11:49202187	rs7124497	SBP	DLRS	1.92E-04	1.22E-04
12:133441378	rs4303268	SBP	QLRS	7.20E-01	6.04E-01
15:64315336	rs12442060	SBP	DLRS	6.93E-01	8.14E-01
15:64316458	rs172995	SBP	DLRS	6.30E-01	7.83E-01
18:30053150	rs11081767	DBP	QLRS	6.21E-01	6.10E-01

^†Chr:position: based on human genome build 37.

Regional association plots for loci detected and their corresponding significant/suggestive SNPs are shown in appendix Figures S4 – S15. Specifically, QLRS found four loci associated with SBP and three loci associated with DBP. Similarly, DLRS found three loci associated with SBP and two loci associated with DBP.

The genome-wide significant locus, intergenic on chromosome 4, was represented by five low-frequency variants, three of which reached genome-wide significance based on the Joint 2 DF test. Three other loci with suggestive associations were also represented by low-frequency variants (rs17636599, rs2012485, and rs7124497). While rs17636599 showed no supporting evidence from surrounding SNPs, rs2012485 and rs7124497 are located within 500 kb of two known BP loci identified through a main effect GWAS (Hoffmann et al., 2017): EBF2 and OR4A47-TRIM51GP, respectively. Similarly, a common variant on chromosome 11, rs10047474, is located within the boundaries of locus GALNT18, which was detected in the CHARGE Gene-Lifestyle Interactions Working Group’s study on gene-alcohol interactions as being associated with SBP in African ancestry (Feitosa et al., 2018).The remaining suggestive associations are novel and were observed for common variants located on chromosomes 2 (NCKAP5 locus), chromosome 3 (SLC9A9 and KALRN loci), chromosome 7 (VIPR2 locus), chromosome 8 (near the OPRK1 locus), chromosome 11 (near the GAREM loci), chromosome 12 (CHFR locus), and chromosome 15 (DAPK2 locus).

All of the suggestive/significant associations identified using LRS (Table 4) were strongly driven by qualitative interactions (opposite signs on ${\beta }_G$ and ${\beta }_{GE}$) between the SNP and the LRS with the exception of SNPs rs7001769 and rs7465458, which have non-significant ${\mathit{P}}_{\mathit{G}\mathit{E}}$ > 0.05. Consistent with this result, the addition of the interaction term made significant contributions to the signal for 19 out of 21 suggestive/significant SNPs; only associations for SNPs rs7001769 and rs7465458 attained nearly suggestive P-values for the genetic effect in the absence of the interaction term (Table 5). For the QLRS, five SNPs showed a BP decrease while three SNPs showed a BP increase for each unit increase in the QLRS. For the DLRS, eight SNPs showed a BP decrease in individuals with a favorable lifestyle and a BP increase in those with an unfavorable lifestyle, while five SNPs showed the reverse pattern.

Six of the detected variants (rs880000, rs12442060, rs7465458, rs2012485, rs7124497, rs11081767) showed regulatory functional evidence. SNP rs880000 in the KALRN locus on chromosome 3 is likely to affect binding (RegulomeDB categorical score of 2b), with the histone modification evidence suggesting that it is located in a region of enhancers in diverse tissues including heart and brain. Intergenic SNP rs7465458, mapped near OPRK1 on chromosome 8, had a CADD score of 22.1, greater than the recommended threshold for deleteriousness (>12.37). All six SNPs were in strong LD with at least one other SNP that had a CADD score greater than the threshold for deleteriousness and/or a RegulomeDB categorical score of 2b or 2c (Table S2). Moreover, our interrogation of the GWAS atlas showed that 16 of 21 variants detected by the gene-LRS analysis had nominal associations with BP traits (including SBP, DBP, and pulse pressure) and/or cardiovascular traits (e.g., coronary artery disease, heart rate recovery, angina pectoris, etc.), while four genetic variants had nominal associations with other cardiovascular traits (e.g., non-ischemic cardiomyopathy). Only one genetic variant had no record of an at least nominal association with a BP or cardiovascular trait (Table S3).

Discussion

In this study, we performed gene-lifestyle interactions in BP traits using a lifestyle risk score (LRS) aggregating four lifestyle exposures (smoking status, alcohol intake, education, and physical activity) in 6,254 FHS participants. Through our gene-LRS interaction approach, we identified one genome-wide significant locus and 11 suggestive loci in this moderate sample size. Among these 12 loci, only one was detected by analyses that used single lifestyle exposures. Considering the LRS as both quantitative and dichotomous variables was complementary, identifying distinct panels of loci. Strikingly, association between BP traits and all but one of the detected loci was driven by qualitative interactions (showing genetic and interaction effects in opposite signs). Also, nine of the 12 loci detected by the LRS are novel, not previously known to be associated with BP. These results point to a valuable, and complementary, role for the multiple lifestyle exposure approach for the discovery of BP loci.

Our results suggest that the nature of some genetic effects on BP, whether harmful or protective, depends on an individual’s overall lifestyle exposure profile as represented by the LRS. Despite a modest sample size, we found evidence of cumulative interaction effects on BP, at suggestive or significant levels, for 21 SNPs in 12 loci. The fact that 19 of the 21 SNPs (in 11 loci) were not detected in the gene-single lifestyle exposure interaction analysis highlights the potential for increased yield of association when cumulative interaction effects are modeled. This potential for finding additional BP loci is noteworthy because, despite its much smaller sample size and statistical power, the LRS analysis found loci that was distinct from loci detected by well powered gene-single lifestyle exposure interaction studies with sample sizes > 500,000. A classic illustration of this potential is for the locus in the KALRN genomic region that we found to be associated with SBP when accounting for interactions with the DLRS. The effect of the KALRN locus on SBP is typified by the index SNP rs880000 for which one copy of the G allele in individuals with a favorable lifestyle decreased SBP by 1.6 mmHg, yet increased SBP by 0.8 mmHg in those with unfavorable lifestyles. In previous GWAS, this region was reported to be associated with cardiovascular diseases such as coronary artery disease and ischemic stroke (Boroumand et al., 2014; Li et al., 2017), but not with BP even when gene-single lifestyle interactions are accounted for. It is known that combined effects of multiple healthy lifestyle exposures on cardiovascular health yield additional benefits not obtained from individual lifestyle exposures, lending credence to the idea that ‘the whole is greater than the sum of the parts’ (Egan, 2018; Lloyd-Jones, 2014). Moreover, a high genetic risk for BP and cardiovascular diseases can be offset by a favorable lifestyle defined by multiple lifestyle exposures (Pazoki et al., 2018). Along with loci identified through the current practice of using the single lifestyle exposure in genome-wide gene-environment interaction scans, the use of the LRS approach may identify additional loci not detected by single lifestyle exposures. Therefore, an effective strategy for maximizing BP loci discovery would be to apply the two approaches in concert, using the LRS as a complement to the single lifestyle exposure approach.

Whether qualitative interactions, characterized by opposite signs for the genetic main and interaction effects, are prevalent but not detected due to inadequate power of existing approaches remains a relevant unanswered question (Winkler et al., 2017). Remarkably, the gene-LRS interaction analysis consistently detected only qualitative interactions, whereas gene-single lifestyle exposure interaction analysis detected both qualitative and quantitative interactions. Given the relevance of finding qualitative interactions to inform personalized medicine and identify high-risk groups, it may be useful to further investigate whether using a measure of overall lifestyle exposure such as the LRS in gene-environment interaction GWAS provides an advantage for detecting opposite-effect loci.

All of the suggestive/significant SNPs that gene-LRS interaction analysis detected are intronic or intergenic, with some showing evidence of regulatory function. This suggests that the resulting influence on BP may be through regulatory activities on nearby protein coding genes, many with effects on neurohormonal regulation, which may play important roles in the pathophysiology of hypertension (Beevers, Lip, & O’Brien, 2001). For example, there were four SNPs that were intronic to the KALRN gene on chromosome 3. KALRN encodes a guanine nucleotide exchange factor that inhibits vascular inducible nitric oxide, an important regulator of vascular hypertrophy and tone (Zhang et al., 2003), and pituitary secretion of adrenocorticotropic hormone (ACTH) (Ratovitski et al., 1999). Associations between variants in the KALRN genomic region and cardiovascular diseases, including early-onset coronary artery disease and ischemic stroke, have been reported (Boroumand et al., 2014; Li et al., 2017). KALRN protein has also been identified as a biomarker of endothelial dysfunction in hypertensive patients with albuminaria (de la Cuesta et al., 2017).

Similarly, there were two SNPs intronic within DAPK2, a gene that plays a primary role in apoptosis. DAPK2 has been shown to reduce blood pressure in rat models by mediating the mTOR signaling pathway, which prevents oxidative stress and apoptosis in hypertensive disorder complicated pregnancy (Wang et al., 2019). Approximately 100 kb upstream of this region is the sorting nexin 1 gene (SNX1). Sorting nexins play a major role in the regulation of trafficking and signal transduction of G-protein coupled receptors (including the renal dopamine receptor, D₅R), a key determinant of water, electrolyte, and BP homeostasis (Yang et al., 2014). An intronic SNP, rs3793217 identifies the gene VIPR2, which encodes a receptor for vasoactive intestinal peptide (VIP), a neurotransmitter and neuromodulator expressed in nearly all tissues (Asnicar et al., 2002), and with known cardiovascular effects including coronary vasodilation, lowering of arterial BP, and regulation of circadian rhythmicity in the heart (Henning & Sawmiller, 2001). Based on evidence from studies of VIPR2 knockout mice, VIPR2 increases insulin sensitivity and regulates circadian rhythm and immune functions (Harmar et al., 2002).

The only locus that had a genome-wide significant association with BP in this study, represented by the index SNP rs11131920, appears to be located in a genomic wilderness, being ∼2 MB away from the nearest mapped gene AGA. Interestingly, this locus has been reported to be associated with the use of several treatment medications (appendix Figures S16), particularly, for hypertension (eprosartan, an angiotensin II receptor blocker; P-value = 4.7 × 10⁻⁸) and for pulmonary arterial hypertension and erectile dysfunction (adalafil, a vasodilator; P-value = 1.5 × 10⁻⁴) (Churchhouse & Neale, 2017). Nevertheless, this genome-wide significant association with BP requires further investigation.

Despite our innovative application of the LRS, there are some limitations. First, our approach to computing the LRS, while simple, is difficult to interpret. This is because it ignores features like correlations among the individual lifestyle exposures (e.g., heavy drinking and current smoking tend to co-occur) and different magnitude of effects on BP traits. Availability of a richer data resource with larger sample size would allow differences in magnitude of effects of component variables to be addressed by deriving the LRS analogous to polygenic risk scores, i.e., the LRS would be computed as a weighted sum of lifestyle exposures, using beta coefficients as weights for each exposure from BP regression on multiple lifestyle exposures (Kim et al., 2019; Park, Tao, Meeker, Harlow, & Mukherjee, 2014). Second, the LRS identified different set of loci depending on whether it was quantitative or dichotomous, which may be due to statistical power. In general, our approach with LRS seems promising and worth pursuing with larger sample sizes and in large consortium settings.

In summary, by considering lifestyle risk score aggregating multiple lifestyle exposures in gene-lifestyle interactions, our limited proof-of-concept investigation appears to have identified several novel and biologically plausible loci. The genetic effects of these loci on BP may have been modulated by the combined effect of multiple lifestyle exposures. Our findings suggest that the LRS approach can contribute to novel discoveries over and above those from single lifestyle exposure analysis, and can complement traditional approaches that use single lifestyle exposures.