A Blood Pressure Genetic Risk Score Is a Significant Predictor of Incident Cardiovascular Events in 32 669 IndividualsNovelty and Significance
Recent genome-wide association studies have identified genetic variants associated with blood pressure (BP). We investigated whether genetic risk scores (GRSs) constructed of these variants would predict incident cardiovascular disease (CVD) events. We genotyped 32 common single nucleotide polymorphisms in several Finnish cohorts, with up to 32 669 individuals after exclusion of prevalent CVD cases. The median follow-up was 9.8 years, during which 2295 incident CVD events occurred. We created GRSs separately for systolic BP and diastolic BP by multiplying the risk allele count of each single nucleotide polymorphism by the effect size estimated in published genome-wide association studies. We performed Cox regression analyses with and without adjustment for clinical factors, including BP at baseline in each cohort. The results were combined by inverse variance–weighted fixed-effects meta-analysis. The GRSs were strongly associated with systolic BP and diastolic BP, and baseline hypertension (all P<10−62). Hazard ratios comparing the highest quintiles of systolic BP and diastolic BP GRSs with the lowest quintiles after adjustment for age, age squared, and sex were 1.25 (1.07–1.46; P=0.006) and 1.23 (1.05–1.43; P=0.01), respectively, for incident coronary heart disease; 1.24 (1.01–1.53; P=0.04) and 1.35 (1.09–1.66; P=0.005), respectively, for incident stroke; and 1.23 (1.08–1.40; P=2×10−6) and 1.26 (1.11–1.44; P=5×10−4), respectively, for composite CVD. In conclusion, BP findings from genome-wide association studies are strongly replicated. GRSs comprising bona fide BP-single nucleotide polymorphisms predicted CVD risk, consistent with a lifelong effect on BP of these variants collectively.
See Editorial Commentary, pp 961–963
Elevated blood pressure (BP) is a strong, independent, and modifiable risk factor for stroke and heart disease.1,2 BP is a heritable trait with estimated heritability of 0.4–0.5,3 and recent well-powered genome-wide association studies (GWAS) have identified several genetic loci that are associated with systolic BP (SBP), diastolic BP (DBP), or commonly both.4–8 Although the variants have modest effects on BP, their presence may act over the entire life course, and therefore lead to substantial increases in risk of cardiovascular and cerebrovascular disease. For example, it was recently found that common genetic variants are associated with preclinical BP traits even in childhood.9 The intraindividual and measurement variability of BP is high,10 and therefore several measurements are optimally needed over time to reliably determine a person’s BP level. In principle, genetic background is stable and could, in borderline cases, help clinicians decide whether BP treatment is needed or whether to alter the intensity of BP treatment.
We genotyped 32 genetic variants that have been previously reported to be associated with BP at genome-wide significance and investigated whether genetic risk scores (GRSs) constructed of these variants would be significant predictors of incident cardiovascular (CVD) events in prospective, population-based cohorts from Finland.
An expanded description of the Methods section is available in the online-only Data Supplement.
FINRISK surveys are cross-sectional, population-based studies conducted every 5 years since 1972 to monitor the risk of chronic diseases. For each survey, a representative random sample was selected from 25- to 74-year-old inhabitants of different regions in Finland. The survey included a questionnaire and a clinical examination, at which a blood sample was drawn, with linkage to national registers of cardiovascular and other health outcomes. The study protocol has been described elsewhere.11 Study participants were followed up through December 31, 2010. The current study included eligible individuals from FINRISK surveys conducted in 1992, 1997, 2002, and 2007 (total n=27 838).
Health 2000 was based on a stratified 2-stage cluster sampling from the National Population Register to represent the total Finnish population aged ≥30 years. The Mini-Finland Health Survey was originally conducted between 1978 and 1980 in similar manner to the Health 2000 Study. Of the Mini-Finland participants, 985 living in 7 large cities participated in a follow-up study in 2001, at which DNA was collected. Health 2000 and Mini-Finland cohorts were analyzed pooled, adjusting for study cohort. The survey included an interview about medical history, health-related lifestyle habits, and a clinical examination, at which a blood sample was drawn. A detailed description of the study protocol is available at: http://www.terveys2000.fi/doc/methodologyrep.pdf. Study participants were followed up through December 31, 2010. After restricting the study to participants aged ≤80 years at baseline, there were 6731 individuals eligible for the current study.
The Helsinki Birth Cohort Study (HBCS) is composed of 8760 individuals born between the years 1934 and 1944 in 1 of the 2 main maternity hospitals in Helsinki, Finland. Between 2001 and 2004, a randomly selected sample of 928 males and 1075 females participated in a clinical follow-up study with a focus on cardiovascular, metabolic, and reproductive health; cognitive function; and depressive symptoms. The participants were followed up through December 31, 2010, and 1676 participants were eligible for the present study.12
The Oulu Project Elucidating Risk of Atherosclerosis (OPERA) is a population-based, epidemiological study examining risk factors and disease end points of atherosclerotic cardiovascular diseases (CVD). The hypertensive cohort (cases) consisted of 600 subjects (300 men and 300 women, aged 40–59 years at the time of selection) from the town of Oulu randomly selected from the national register of medication reimbursements for moderate or severe hypertension (HTN). For each year of birth (1931–1950), 15 hypertensive men and 15 hypertensive women were selected. For each hypertensive subject, an age- and sex-matched control was randomly selected from all inhabitants of Oulu, excluding subjects with reimbursement for HTN medication.13 Study participants were followed up through December 31, 2009. For the present study, 1000 participants were eligible. BP measurement methods for each of the cohorts are described in the online-only Data Supplement.
During follow-up, hospitalization and mortality data were obtained from the Finnish National Hospital Discharge Register and the National Causes-of-Death Register. These registers cover all cardiovascular events that have led either to hospitalization or death in Finland. The cardiovascular diagnoses in these registers have been validated.14,15 Coronary heart disease (CHD) was defined as nonfatal myocardial infarction, unstable angina pectoris, coronary revascularization (coronary artery bypass graft surgery or percutaneous transluminal coronary angioplasty), or death attributable to CHD. CVD included CHD and ischemic stroke events. The clinical outcomes were linked to study subjects using their unique national social security ID, which is assigned to every permanent resident of Finland. Events before baseline were traced back to 1970, when computerized records first became available in Finland. In the OPERA study, however, register data were available only from the baseline onwards, and events before baseline were determined by a detailed interview of a physician. All study subjects provided written informed consent. The local institutional ethical review boards approved each study.
Single Nucleotide Polymorphism Selection, Genotyping, and GRSs
We selected 32 single nucleotide polymorphisms (SNPs; Table S1 in the online-only Data Supplement), which have been associated with SBP or DBP in GWAS.4–8 Details of the genotyping are provided in the online-only Data Supplement. We calculated the GRSs using the reported effect sizes from the reference studies as weights per copy of the coded allele for each individual SNP. Different SNPs thus contribute different weights, as opposed to an alternate approach in which no weighting of effects is used, and each SNP allele counts equally in the score. The coded allele is the allele coded 0, 1, or 2 according to the number of copies of the allele. A single unit increase in the GRS corresponds to a 1-mm Hg increase in the predicted SBP or DBP, as a result of the aggregated predicted effects of all 32 SNPs. Missing genotype data for each SNP were imputed using the average coded allele frequency within each study cohort. However, if >60% of the SNP genotypes were missing for a given individual, the GRS was set as missing for that individual. Two GRSs were calculated, 1 for SBP and 1 for DBP.
Associations of the GRSs and, as a secondary analysis, individual SNPs with SBP or DBP with imputation for use of antihypertensive therapy (+15 mm Hg for SBP; +10 mm Hg for DBP as an estimate of the expected BP change off antihypertensive therapy)16 was performed using linear regression, adjusting for baseline age and its square, sex, and geographic region (eastern versus western Finland).The analysis was repeated with further adjustment for body mass index, alcohol consumption (g/wk), and leisure time physical activity (moderate to high versus low). Association of the GRSs and individual SNPs with baseline HTN (defined as SBP≥140 mm Hg or DBP≥90 mm Hg or use of antihypertensive therapy) was analyzed using logistic regression adjusted for the same covariates as above.
Associations of the GRSs and secondarily individual SNPs with incident CVD outcomes were analyzed using Cox proportional hazards regression with age as the time scale. All models were stratified by sex and adjusted for geographic region. As a secondary analysis, we included also prevalent CVD cases, that is, using the first event since birth. In these analyses, we used time from birth as the time scale except in HBCS, which used elapsed time from the first date of eligibility, January 1, 1970. Power analysis for detecting a single SNP effect in the Cox regression for incident CVD is provided in Figure S1.
The incident outcome associations were then analyzed using 2 further adjustments. First, we adjusted for the following Framingham Risk Score risk factors: total cholesterol, high-density lipoprotein cholesterol, current smoking and baseline diabetes mellitus, plus lipid-lowering treatment. The last covariate is not part of the Framingham Risk Score, but it was significant in most models, improving model fit and validity of the proportional hazards assumption. Second, we adjusted additionally for SBP or DBP (in the case of the DBP GRS) and for antihypertensive treatment. In addition to the analyses of the primary CVD end point, we performed secondary Cox regression analyses using ischemic stroke and CHD events as the outcomes with the same covariate adjustments as in the primary analyses.
To examine the possibility that the GRS derived for SBP or DBP would not be the optimal GRS for predicting incident cardiovascular events, we derived another GRS specifically for CVD, using the Cox regression results of the individual SNPs with incident CVD in the present studies. This CVD GRS used betas or hazard ratios from the meta-analyses (Table S3) as weights for the risk allele counts.
We confirmed that Cox proportional hazards assumptions were met using scaled Schoenfeld residuals (R function cox.zph). The results from individual studies were combined using inverse variance–weighted fixed-effect meta-analysis, checking for heterogeneity using the I2 measure; I2>0.5 is considered evidence for significant heterogeneity.17 This limit was not exceeded in any of the analyses, beyond chance expectation.
Using the Health 2000 and FINRISK 92 and 97 cohorts, which have 10-year follow-up, we further assessed the utility of the GRSs for 10-year CVD risk prediction by estimating the net reclassification improvement (NRI), clinical NRI for prospective data,18 integrated discrimination improvement,19 and explained relative risk for the GRSs. The statistical significance of the area under the receiver operating characteristic (ROC) curve change between models with and without GRSs was tested with the correlated C-index approach.20 Model calibration was assessed with the Hosmer-Lemeshow goodness-of-fit test.21 All statistical analyses were performed using R version 2.15.22 In general, we considered 2-sided P<0.05 statistically significant. For the association tests of individual SNPs with BP/HTN and cardiovascular events, we used Bonferroni correction to account for 32 independent tests.
The baseline characteristics of the study cohorts are given in Table 1. Individually, 23 of the 32 SNPs were associated with SBP or DBP in the same direction, as previously reported, after accounting for multiple testing (P<0.0016 = P<0.05/32; Table S1). Directions of effect were consistent for all SNPs but 2. Effects on SBP or DBP were highly correlated in the Finnish studies with the original estimated effects reported in the literature (SBP: r=0.75; DBP: r=0.71). GRSs were strong predictors for SBP, DBP, and HTN (Table 2; all P<10−62). The average proportion of variance in SBP or DBP explained by each respective score was generally greater than that estimated previously8: 1.20% in SBP and 1.18% in DBP using weighted averages across all cohorts. The results for the individual cohorts are provided in Table S4.
After excluding 1284 individuals with prevalent CVD at baseline, 2295 incident CVD events occurred during a median follow-up time of 9.8 (interquartile range, 5.1) years. In total, the study participants contributed 347 955 person-years of follow-up. GRSs showed significant, independent, and roughly linear associations with CVD risk (Tables 3 and 4; Figure 1). As expected, the observed effects on SBP or DBP of a predicted 1-mm Hg increase in GRS, based on the previously published per-SNP effect estimates, were 1.1 and 1.0 mm Hg, for the SBP and DBP GRSs, respectively (Table 2). A 4% increased hazard for CHD, stroke, or CVD was observed for each predicted 1-mm Hg increase in SBP GRS, and 6% to 8% for each predicted 1-mm Hg increase in DBP GRS, in models adjusting for non-BP clinical risk factors. After further adjustment for antihypertensive treatment and baseline SBP or DBP for their respective GRSs, the hazard ratios were reduced only slightly and for the most part remained statistically significant (Table 3). In models, adjusting for age, age squared, and sex, increasing quintile of SBP or DBP GRS was associated with roughly linear increases in BP, HTN prevalence, and risk of incident CVD, with or without inclusion of prevalent cases (Table 4). For example, the highest compared with lowest quintile of SBP GRS was associated with a hazard ratio for CVD of 1.30 (95% confidence interval, 1.17–1.45; P=1×10−6), including prevalent and incident events, and 1.23 (95% confidence interval, 1.08–1.40; P=2×10−6), including only incident events, and the highest quintile of DBP GRS was associated with a hazard ratio of 1.30 (95% confidence interval, 1.17–1.45; P=2×10−6) and 1.26 (95% confidence interval, 1.11–1.44; P=5×10−4), respectively (Table 4). No individual SNPs were statistically significantly associated with CVD risk after correcting for 32 tests, although some were nominally associated (P<0.05; Table S3).
The associations of a CVD GRS, created using CVD effects estimated in the Finnish cohorts, with cardiovascular events were stronger than the associations with the BP GRSs with less attenuation after adjustment for Framingham Risk Score components (hazard ratio=2.40, P=6×10−5 for incident CHD; hazard ratio=2.07, P=0.013 for incident stroke; and hazard ratio=2.40, P=2×10−6 for incident CVD; Table S5). However, because the CVD GRS was derived from the same studies, in which its association with CVD events was tested, the strength of the association might have been overestimated.
Reclassification analyses in Health 2000 and FINRISK 1992 and 1997 showed that the SBP GRS did not improve CVD risk discrimination over and above the standard Framingham Risk Score (which includes SBP), when assessed using ROC curves and C-statistics, or integrated discrimination improvement (Tables S6 and S7). There was no net improvement in NRI either (Tables S6 and S7). However, the clinical NRI, that is, reclassification in the intermediate risk group of 5% to 20%, was 3.8% and statistically significant (P=5×10−5). Model calibration was good in both models with and without the GRSs (Figure S2). We found no significant interaction of sex, baseline age, body mass index, or antihypertensive treatment with the GRS effect on risk of incident CVD.
We tested previously established SNPs associated with BP in a large collection of population-based studies in Finland. We replicated 23 SNPs at a stringent Bonferroni-corrected significance threshold and found directional consistency for 30 of 32 SNPs. GRSs weighted according to previously reported effect estimates were highly associated with BP and HTN. An important novel finding of our study was that both SBP- and DBP-based GRSs were strongly associated with risk of incident CVD, among those free of CVD at baseline. These associations were largely independent of standard CVD risk factors, including BP measured at baseline.
The field of genetic associations has historically been riddled with irreproducible results, largely attributable to overly permissive significance thresholds and inadequate power from limited sample sizes.23 The widespread adoption of stringent P value thresholds and the development of genotyping platforms allowing the efficient genotyping of scores to millions of variants in tens of thousands of individuals have enabled identification of reproducible associations. We have replicated many of the BP variants identified in GWAS meta-analyses, some of which included 6 times the sample size examined here.
Our findings are consistent with a causal effect of BP on CVD. This is well accepted, given the broad epidemiological support for this relationship, the existence of Mendelian syndromes of HTN and premature CVD, and the modification of risk by lowering BP through behavioral or pharmacological means.24 Most of the SNPs that have been identified lie in chromosomal regions with a heterogeneous set of genes without a single dominant pathway apparent. Despite these heterogeneous effects, when the weak effects of the individual SNPs are considered jointly, a strong and consistent effect on cardiovascular risk is observed.
BP is a highly variable, dynamic measure, with minute-to-minute variation influenced by activity, posture, mental stress, medication, etc. We used the mean of 2 BP measurements. However, we were limited to the examination of BP at a single time point, which is clearly an imprecise surrogate for the lifetime of exposure to higher BP that contributes to the pathogenesis of CVD and could lead to regression dilution of the true impact of BP and other time-varying factors. This, in fact, highlights the potential value of the study of precisely measured genotypes. Genetic variants are fixed and, when set against the background of a large dynamic range, may capture a fixed component to lifetime BP exposure. Thus, small genetic effects on BP may translate into comparatively large effects when compounded over a lifetime. In the present study, we found no net improvement in NRI attributable to SBP or DBP GRS over and above the Framingham Risk Score. It should be noted, however, that the Framingham Risk Score estimates 10-year risk only, because risk factors change over time, whereas for BP GRS, the estimation of lifetime risk might be more appropriate, because genotype is invariant over time.
For decades, investigators have sought to resolve essential HTN into well-defined subtypes that might benefit from specific therapies, such as low-renin, high-aldosterone HTN. However, these have not led to widespread adoption or recommendation for specific therapies by most public health guidelines.24 Given the limited power to detect the weak SNP effects that have so far been found, it is estimated that hundreds of common variants of similarly modest effect will ultimately be found to exist.8 Assuming as yet unidentified effects are as modest as those identified to date, it seems unlikely that common genetic variation will ultimately accomplish this task.
Some SNPs that fall in targets of antihypertensives are potential candidates to modulate the response to antihypertensive therapy. For example, a common BP-associated missense polymorphism lies in ADRB1, which encodes the β1 adrenergic receptor, a target of beta adrenergic receptor antagonists. Whether such BP-associated variants also influence response to antihypertensive therapy awaits results from large clinical trials.
Strengths of the current study include the large sample size, the population-based sampling of the cohorts examined, the measurement of BP precisely and in a uniform manner, the availability of relevant covariates, the large number of CVD outcomes available for prospective analyses, and the single country of origin for all samples. Limitations include the inability to generalize to non-European ancestry groups. We lacked the ability to adjust for time-varying clinical factors that influence BP. The proportion of variation explained by the SNPs remained low, and the level of prediction for events was also relatively small. We lacked power to demonstrate replication even at nominal P<0.05 of some of the modest BP effects reported in previous GWAS.
We found that GRSs comprising 32 SNPs identified in GWAS of BP were strongly associated with risk of incident CVD, even after adjustment for baseline BP. Although the GRSs were not associated with significant reclassification of CVD risk when added to the Framingham Risk Score, a more complete compendium of genetic variation that reproducibly influences BP will ultimately need to be tested.
We thank Annukka M. Lahtinen from the Research Programs Unit, Molecular Medicine and Department of Medicine, University of Helsinki, who participated in the Sequenom genotyping; Antti-Pekka Sarin from the Institute for Molecular Medicine Finland (FIMM), University of Helsinki, who did the genotyping QC; and Dr Merja Santaniemi from the Institute of Clinical Medicine, Department of Medicine, University of Oulu, who helped in coding the follow-up events in the OPERA cohort.
Sources of Funding
FINRISK was mainly funded by the National Institute for Health and Welfare. Additional support was obtained from the Academy of Finland (grant numbers 129 494 and 139 635) to Dr Salomaa. The Health 2000 Study is funded by the National Institute for Health and Welfare (THL), the Finnish Center for Pensions (ETK), The Social Insurance Institution of Finland (KELA), The Local Government Pensions Institution (KEVA), and other organizations listed on the website of the survey (http://www.terveys2000.fi). The OPERA study was supported by the Finnish Foundation for Cardiovascular Research. The Helsinki Birth Cohort Study has been supported by grants to Dr Eriksson from the Academy of Finland (Grant numbers 129 255, 126 775, 135 072), the Finnish Diabetes Research Society, Folkhälsan Research Foundation, Novo Nordisk Foundation, Finska Läkaresällskapet, Signe and Ane Gyllenberg Foundation, Ahokas Foundation, and Juho Vainio Foundation. Dr Kontula was supported by the Sigrid Juselius Foundation and the Finnish Foundation for Cardiovascular Research. Dr Newton-Cheh and the SNP genotyping were supported by the Burroughs Wellcome Fund and the National Institutes of Health (HL098283).
Dr Newton-Cheh reports that he is a member of a scientific advisory board for hypertension and heart failure at Merck. The other authors have no conflicts to report.
The online-only Data Supplement is available with this article at http://hyper.ahajournals.org/lookup/suppl/doi:10.1161/HYPERTENSIONAHA.111.00649/-/DC1.
- Received November 14, 2012.
- Revision received January 8, 2013.
- Accepted February 19, 2013.
- © 2013 American Heart Association, Inc.
- MacMahon S,
- Peto R,
- Cutler J,
- Collins R,
- Sorlie P,
- Neaton J,
- Abbott R,
- Godwin J,
- Dyer A,
- Stamler J
- Collins R,
- Peto R,
- MacMahon S,
- Hebert P,
- Fiebach NH,
- Eberlein KA,
- Godwin J,
- Qizilbash N,
- Taylor JO,
- Hennekens CH
- Levy D,
- DeStefano AL,
- Larson MG,
- O’Donnell CJ,
- Lifton RP,
- Gavras H,
- Cupples LA,
- Myers RH
- Johnson AD,
- Newton-Cheh C,
- Chasman DI,
- et al
- Oikonen M,
- Tikkanen E,
- Juhola J,
- et al
- Vartiainen E,
- Laatikainen T,
- Peltonen M,
- Juolevi A,
- Männistö S,
- Sundvall J,
- Jousilahti P,
- Salomaa V,
- Valsta L,
- Puska P
- Kiema TR,
- Kauma H,
- Rantala AO,
- Lilja M,
- Reunanen A,
- Kesäniemi YA,
- Savolainen MJ
- 22.↵R Development Core Team R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, 2011.
- Chobanian AV,
- Bakris GL,
- Black HR,
- Cushman WC,
- Green LA,
- Izzo JL Jr.,
- Jones DW,
- Materson BJ,
- Oparil S,
- Wright JT Jr.,
- Roccella EJ
Novelty and Significance
What Is New?
Recent genome-wide association studies have identified several genetic variants associated with blood pressure (BP).
The predictive power of BP-associated genetic variants for incident cardiovascular disease events has not been established.
What Is Relevant?
Genetic risk scores comprising 32 single nucleotide polymorphisms identified in genome-wide association studies of BP were strongly associated with risk of incident cardiovascular disease, even after adjustment for baseline BP and antihypertensive treatment.
A more complete compendium of genetic variation that reproducibly influences BP will ultimately need to be tested.
The association of blood pressure variants with cardiovascular disease risk is consistent with a lifelong effect on BP of genetic variants associated with BP and a causal effect of BP on cardiovascular disease risk.