Association between a Literature-Based Genetic Risk Score and Cardiovascular Events in 19,313 Women

Abstract

Context

While multiple genetic markers associated with cardiovascular disease (CVD) have been identified by genome wide association studies, their aggregate effect on risk beyond traditional factors is uncertain, particularly among women.

Objective

To test the predictive ability of a literature-based genetic risk score for CVD.

Design, Setting, and Participants

Prospective cohort of 19,313 initially healthy white women in the Women’s Genome Health Study followed over a median of 12.3 years. Genetic risk scores were constructed from the National Human Genome Research Institute’s catalog of genome-wide association study results published between 2005 and June 2009.

Main Outcome Measure

Incident myocardial infarction, stroke, arterial revascularization, and cardiovascular death.

Results

We identified 101 single nucleotide polymorphisms reported to be associated with CVD or at least one intermediate CVD phenotype at a published p-value less than 10^-7 and added the risk alleles to create a genetic risk score. During follow-up 777 CVD events occurred (199 myocardial infarctions, 203 strokes, 63 cardiovascular deaths, 312 revascularizations). After adjustment for age, the genetic risk score had a hazard ratio for CVD of 1.015 per risk allele (95% confidence interval: 1.00-1.03, p value 0.006). This corresponds to an absolute CVD risk of 3% over 10 years in the lowest tertile of genetic risk (73-99 risk alleles) and 3.7% in the highest tertile (106-125 risk alleles). However, after adjustment for traditional factors, the genetic risk score did not improve discrimination or reclassification (change in c-index from ATPIII 0.000; NRI 0.5%, p=0.24), nor was it associated with CVD risk (ATPIII adjusted HR per allele 1.00, 95% CI 0.99-1.01). In contrast, self-reported family history remained significantly associated with CVD in multivariable models.

Conclusions

After adjustment for traditional cardiovascular risk factors, a genetic risk score comprising 101 single nucleotide polymorphisms was not significantly associated with the incidence of total cardiovascular disease.

Introduction

Risk prediction is a central part of cardiovascular disease prevention and refining prediction strategies remains important for targeting treatment recommendations. One area of potential improvement has been the discovery of genetic markers for cardiovascular disease as well as intermediate phenotypes such as cholesterol and blood pressure. Recent efforts using genome-wide association studies have greatly expanded the discovery of genetic markers associated with cardiovascular disease.

To date, however, the utility of single genetic markers to improve cardiovascular risk prediction has shown mixed results, even for the most promising marker, located in the 9p21 region.^1-3 To combine the relatively small effects of individual genes and to better capture the complex relationship between genetics and cardiovascular disease, the use of a multi-locus genetic risk score (GRS) has been proposed.⁴ One such score developed by Kathiresan and colleagues,⁵ included nine genetic markers associated with increased lipid levels, but showed no improvement in discrimination and only a slight improvement in reclassification. In large part, however, the predictive abilities of recently discovered genetic markers have been untested.⁶ In particular, there has been no evaluation of a literature-based cardiovascular disease genetic risk score, a possibility that is facilitated by the online catalog, maintained by the National Human Genome Research Institute (NHGRI), of all genetic markers identified through genome wide association studies.⁷

We constructed two genetic risk scores based on a comprehensive literature-based selection of genetic markers known to be associated with either cardiovascular disease or an intermediate phenotype selected from the NHGRI catalog. The scores were then tested to assess their predictive ability in the Women’s Genome Health Study. We additionally assessed the predictive ability of genetic information alone, as well as in combination with known cardiovascular risk factors and compared the genetic information to self-reported family history.

Methods

Genetic Marker Selection

The single nucleotide polymorphisms (SNPs) which make up the genetic risk scores tested were selected using the online catalog of genome wide association studies from the National Human Genome Research Institute as of June 5, 2009.⁷ In brief, the catalog is a curated and regularly updated list of all published association between SNPs and human disease phenotypes with a p-value less than 10^-5 from studies which examined at least 100,000 SNPs. From this list we selected all SNPs with published association with either cardiovascular disease (myocardial infarction, stroke, coronary disease, and/or cardiovascular death) or an intermediate phenotype (total cholesterol, high density lipoprotein cholesterol, low density lipoprotein cholesterol, triglycerides, blood pressure, diabetes, hemoglobin A1c or fasting blood glucose, and high sensitivity C-reactive protein) where the p-value was less than 10^-7. The original reports for all identified SNPs were used to confirm the published risk allele (the allele which increased the level or probability) for the phenotype. The published risk allele was designated the cardiovascular risk allele for all phenotypes except HDL, for which the lowering allele was designated. To limit our results to independent effects, SNPs in each chromosome were pruned to ensure a linkage disequilibrium r² less than 0.5 using the pairwise pruning function in Plink (http://pngu.mgh.harvard.edu/purcell/plink/).⁸

Two genetic risk scores were constructed on an a priori basis. The first genetic risk score was the sum of all cardiovascular risk alleles from all SNPs, both those associated with CVD and those associated with risk factors. SNPs affecting more than one phenotype were only included once. The second genetic risk score was created by limiting the list to only SNPs with a published association with CVD before pruning and then adding the number of risk alleles. We assumed additive and independent effects for each risk allele; for both scores simple counts of the total number of risk alleles were used rather than weighting by the effect of each SNP. We chose to use an unweighted approach because the current literature was insufficient to provide stable estimates for each effect, all anticipated effects based on published data were of small magnitude, and using weights from the Women’s Genome Health Study data itself would introduce bias into the results.

Study Population

The Women’s Genome Health Study ⁹ is an ongoing prospective cohort, derived from the Women’s Health Study ¹⁰. It includes over 25,000 initially healthy female health professionals who provided a baseline blood sample as well as extensive survey data. For this study we limited our analysis to participants for whom complete data was available for both the traditional risk factors and for the genetic risk scores. We further restricted to the self-reported white participants to both avoid population stratification and because many of the published genetic associations have been thus far only explored in white populations. These restrictions resulted in 19,313 women for the testing of the genetic scores. All participants provided consent for blood-based analyses and long-term follow-up and the study was approved by the institutional review board of the Brigham and Women’s Hospital (Boston, Massachusetts).

Information on age, race, smoking status, blood pressure, hypertension treatment, diabetes, and parental history of myocardial infarction (MI) before 60 years was collected by questionnaire at the beginning of the study. Plasma biomarkers for total cholesterol, high and low density lipoprotein cholesterol, triglycerides, hemoglobin A1c, and high-sensitivity C-reactive protein (hsCRP) were analyzed in a core laboratory facility, certified by the National Heart, Lung and Blood Institute/Centers for Disease Control and Prevention Lipid Standardization Program.

Genetic information was collected using the Illumina HumanHap300 Duo “+” platform which contains both a standard panel of approximately 317,000 single nucleotide polymorphisms (SNPs) for capturing variation among individuals with European ancestry as well as approximately 45,000 SNPs selected specifically for their potential relationship with cardiovascular and other disease. SNPs defining the APOE alleles were available using an oligonucleotide ligation procedure.^{11, 12} In order to use published SNPs that were not directly genotyped, we used the MACH 1.0.16 program (http://www.sph.umich.edu/csg/abecasis/mach/index.html) and information from the HapMap ¹³ to impute additional genotypes. The MACH program has been shown to have high accuracy ¹⁴ and only SNPs with an estimated squared correlation between the imputed and true genotype greater than 0.3 were included, which provides high sensitivity and specificity.¹⁵ Of the 101 SNPs selected, 46 were measured directly and 55 were imputed, with a minimum R squared of 0.6 and 80% above 0.8. The estimated maximum likelihood number of alleles was used in the risk score.

Participants were followed for a median of 12.3 years for incident MI, ischemic stroke, coronary revascularization, and cardiovascular deaths which were combined to calculate total CVD. All endpoints were adjudicated using additional medical records.

Statistical Methods

Cox proportional hazards models were used to generate estimates of predicted risk using a base model with and without each genetic risk score. The base models examined were age alone, covariates from the Third Report of the National Cholesterol Education Program (NCEP) Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults (ATP III) risk score based on the Framingham cohort as well as history of diabetes (noted as a high-risk equivalent)¹⁶, and covariates from the Reynolds Risk Score, a previously published model which includes hemoglobin A1c and C-reactive protein as well as data on family history.¹⁷ The estimated predicted risks were then compared using Harrell’s c-index ¹⁸ to examine discrimination, which measures whether a prediction method ranks cases higher than non-cases, and the Hosmer-Lemeshow goodness-of-fit test ¹⁹ to examine calibration, which measures how well the predicted number of events match up with the observed number of events. Reclassification was assessed by comparing the predicted 10-year risk for each pair of models (base model alone vs. base model plus genetic score) across the categories of less than 5% risk, 5% to less than 10%, 10% to less than 20%, and 20% or higher risk. From the resulting reclassification table, we computed the reclassification calibration statistic²⁰, which assesses the match between predicted and observed event rates for each model in each division of the table with lower values and higher p-values suggesting better fit. Reclassification calibration statistics cannot be directly compared across different models, but large differences between models can suggest differences in fit. We also computed the Net Reclassification Improvement (NRI) ²¹ for the women with complete 10 year follow-up, which examines whether the addition of the genetic risk score moves cases to higher risk categories more often than lower risk and controls to lower risk categories more often than higher risk; the null value is 0%, corresponding to equal movement in the correct and incorrect directions.

Statistical significance was considered to be met with a p-value less than 0.05 and all testing was 2-sided. All statistical analysis was done using R, version 2.6. Using the distribution of the 101 SNP genetic risk score in our data, we had 90% power to detect a 10-year odds ratio per allele as low as 1.0124.

Results

Using the NHGRI catalog, we identified 157 SNPs with a published risk allele and a p-value less than or equal to 10^-7 for the association with cardiovascular disease or an intermediate phenotype which were able to be matched in our geneotyped or imputed data. An additional 5 SNPs were not able to be matched (rs17465637²² in MIA3 gene region, rs28927680²³ in the APOA1/C3/A4/A5 region, rs3812316²⁴ and rs326²⁴ in the MLXIPL gene region, and rs4712524²⁵ in the KCNQ1 gene region). After pruning to eliminate correlated SNPs in high linkage disequilibrium, the remaining 101 SNPs were used in the construction of the primary genetic risk score. The second score, limited to SNPs with a published association with incident CVD, included 12 SNPs after pruning.

The resulting genetic scores were evaluated in the 19,313 white participants from the WGHS. At baseline the participants had a median age of 52.8 years (25^th -75^th percentile: 48.9 - 58.9), a median systolic blood pressure of 125 mm Hg (25^th -75^th percentile: 115 - 135), a median total cholesterol of 208 mg/dL (25^th -75^th percentile:184 - 235), a median high density lipoprotein cholesterol of 52 mg\dL (25^th -75^th percentile: 43.3 - 62.5), and a median high-sensitivity C-reactive protein of 2 mg/dL (25^th -75^th percentile: 0.8 - 4.3). Also at baseline, 12% (2248) were current smokers and 2% (479) were diabetic. In the diabetics, the median hemoglobin A1c level was 6.9% (25^th -75^th percentile: 5.9 - 8.3). Thirteen percent of the women (2499) reported a parental history of MI before 60 years. Over the follow-up period (median 12.3 years), 777 incident cardiovascular events (199 myocardial infarctions, 203 strokes, 63 cardiovascular deaths, 312 revascularizations) were reported by the study participants and confirmed by the endpoints committee (634 in the first 10 years).

The 101 SNPs used in the genetic risk score are shown in eTable1 arranged by the category of the phenotype for the published association. The 12 SNPs used for the score based only on SNPs known to be associated with cardiovascular disease are listed in the phenotype category “Cardiovascular Disease”. Each SNP was tested for association with the previously published phenotype as well as for association with incident CVD in the WGHS. These results, along the candidate gene, the published cardiovascular risk allele, and the frequency of the risk allele in the WGHS, are also included in Table 1. Of the101 SNPs, 72 replicated the published phenotype association in the WGHS with a p-value less than 0.05 and 5 were significantly associated with incident CVD (rs17249754 in the ATP2B1 gene region, rs1333049 in the chromosome 9p21.3 region, rs10830963 in the MTNR1B gene region, rs4607103 in the ADAMTS9 gene region, and rs1883025 in the ABCA1 gene region). Only rs1333049 in the chromosome 9p21.3 region has a previously published genome-wide association with cardiovascular disease.

Table 1.

Association of Genetic Risk Score and Family History with Cardiovascular Disease

Base Model	101 SNP Genetic Risk Score a		12 SNP Genetic Risk Score b		Family History of Premature MI
	HR per allele (95% CI)	p-value	HR per allele (95% CI)	p-value	HR (95% CI)	p-value
Age	1.02 (1.00, 1.03)	0.006	1.05 (1.01, 1.09)	0.014	1.67 (1.39, 1.03)	<0.001
ATP III covariates c	1.00 (0.99, 1.01)	0.63	1.04 (1.00, 1.08)	0.052	1.57 (1.31, 1.89)	<0.001
Reynolds covariates d	1.00 (0.99, 1.01)	0.76	1.04 (1.00, 1.07)	0.06	-	-

^aIncludes SNPs (single nucleotide polymorphisms) associated with incident cardiovascular disease and intermediate phenotypes

^bIncludes only SNPs associated with incident cardiovascular disease

^cAge, systolic blood pressure, hypertensive medication use, smoking, diabetes, total and high density lipoprotein cholesterol

^dAge, systolic blood pressure, smoking, diabetes, total and high density lipoprotein cholesterol, c-reactive protein, family history of premature myocardial infarction

Among the 19,313 WGHS participants, the mean score (or number of risk alleles) using the 101 SNPs was 102.1, with a standard deviation of 6.4 and a range of 73 to 125. The mean score using the 12 SNPs was 10.7, with a standard deviation of 1.9 and a range of 4 to 19. As anticipated, the 101 SNP score was positively correlated with total cholesterol, systolic blood pressure, and C-reactive protein, and negatively associated with high density lipoprotein (eTable2). The 12 SNP score was also positively correlated with total cholesterol, but the relationship was sharply attenuated when the one SNP with a published association with cholesterol levels (rs599839 in the CELSR2/PSRC1/SORT1 region) was removed. The odds of a family history of premature MI also increased with increasing scores, with an odds ratio of 1.01 per allele for the 101 SNP score and 1.04 per allele for the 12 SNP score, both with p-values less than 0.001.

The figure shows the unadjusted survival curves by tertile of each genetic risk score and for family history (top row), as well as the distribution of risk alleles by event status at 10 years of follow-up for the genetic risk scores (bottom row). While there is a trend towards increasing risk with greater number of risk alleles for both scores, only the highest tertile of the 101 SNP score had a significant hazard ratio (HR 1.22 compared to the lowest risk group, p = 0.027). This corresponds to an absolute CVD risk of 3% over 10 years in the lowest tertile of genetic risk (73-99 risk alleles) and 3.7% in the highest tertile (106-125 risk alleles). As suggested by the overlap in the distributions by event status, neither genetic risk score alone had discriminatory capabilities for CVD risk (c-index 0.523 for 101 SNP score and 0.517 for 12 SNP score).

Figure. Cumulative Incidence of Cardiovascular Events by Genetic Risk Score Tertile and Family History and Distribution of Genetic Risk Scores by 10 Year Cardiovascular Event Status.

101 SNP genetic risk score tertile 1: mean 95, range 73-99, tertile 2: mean 102, range 100-105, tertile 3: mean 110, range 106-125; 12 SNP score tertile 1: mean 9, range 4-10, tertile 2: mean 11, range 11-12, tertile 3: mean 14, range 13-19. The y-axis for the distribution graphs is the proportion of the group (either with or without a CVD event at 10 years) with a given risk score; the lines are generated with a Gaussian kernel density smoother.

Both the 101 and 12 SNP genetic risk scores were associated with increased risk of CVD after adjusting for age (Table 1). Specifically, the age-adjusted hazard ratio for CVD per allele for the 101 SNP score was 1.015 (95% confidence interval: 1.00-1.03, p value 0.006) and was 1.05 (95% confidence interval: 1.01-1.09, p value 0.014) for the 12 SNP score. Neither score remained independently associated once the ATP III or Reynolds covariates were adjusted for. The ATP III adjusted HR per allele for the 101 SNP score was 1.00 (95% CI 0.99-1.01) and was 1.04 (95% CI 1.00-1.08) for the 12 SNP score. In contrast, family history of premature MI remained an independent risk factor for incident cardiovascular disease even after adjustment (HR 1.57, 95% CI 1.31-1.89). The effects of the standard risk factors were unaffected by the addition of the genetic markers (models shown in eTables 3 and 4).

All of the models were calibrated with and without with the addition of the genetic risk scores or family history. Neither genetic risk score improved prediction when added the ATP III or Reynolds covariates, as shown in Table 2. Adding the 101 SNP score to the ATP III covariates changed the c-index 0.000 with an NRI of 0.5% (p=0.24), whereas adding the 12 SNP score changed the c-index 0.001 (p=0.12) with an NRI of 0.5% (p=0.59) The 12 SNP risk score and family history did show some improvement in prediction beyond age alone. When the reclassification calibration was examined (Table 3), only family history showed an improvement in fit when added to the base models.

Table 2.

Discrimination and Reclassification after Addition of Genetic Risk Score or Family History to Base Model

Base Model		101 SNP Genetic Risk Score a				12 SNP Genetic Risk Score b				Family History of Premature MI c
		Discrimination		Reclassification		Discrimination		Reclassification		Discrimination		Reclassification
	C Index	C Index	p-valued	NRI	p-valuee	C Index	p-valued	NRI	p-valuee	C Index	p-valued	NRI	p-valuee
Age	0.701	0.704	0.14	1.2	0.13	0.705	0.01	0.6	0.52	0.709	0.013	3.1	0.02
ATP III covariates f	0.793	0.793	0.92	0.5	0.24	0.794	0.12	0.5	0.59	0.796	0.059	1.4	0.28
Reynolds covariates g	0.796	0.796	0.84	0.4	0.21	0.796	0.12	0.8	0.36	-	-	-	-

^aIncludes SNPs (single nucleotide polymorphisms) associated with incident cardiovascular disease and intermediate phenotypes

^bIncludes only SNPs associated with incident cardiovascular disease

^cParental MI (myocardial infarction) before age 60

^dP-value is for comparison with the base model c-index

^eP-value is compared to the null NRI of 0% or equal reclassification correctly and incorrectly

^fAge, systolic blood pressure, hypertensive medication use, smoking, diabetes, total and high density lipoprotein cholesterol

^gAge, systolic blood pressure, smoking, diabetes, total and high density lipoprotein cholesterol, c-reactive protein, family history of premature myocardial infarction

Table 3.

Reclassification Calibration for the Addition of Genetic Risk Score or Family History to Base Model

Base Model	101 SNP Genetic Risk Score a				12 SNP Genetic Risk Score b				Family History of Premature MI c
	Chi2 Base	p-value	Chi2 Plus	p-value	Chi2 Base	p-value	Chi2 Plus	p-value	Chi2 Base	p-value	Chi2 Plus	p-value
Age	9.9	0.13	8.9	0.18	5.9	0.44	4.8	0.56	16.9	0.01	3.6	0.73
ATP III covariates d	11.6	0.009	11.6	0.009	14.7	0.04	14.2	0.05	22.1	0.005	13.3	0.10
Reynolds covariates e	4.1	0.25	4.1	0.25	9.4	0.15	8.4	0.21	-	-	-	-

^aIncludes SNPs (single nucleotide polymorphisms) associated with incident cardiovascular disease and intermediate phenotypes

^bIncludes only SNPs associated with incident cardiovascular disease

^cParental MI (myocardial infarction) before age 60

^dAge, systolic blood pressure, hypertensive medication use, smoking, diabetes, total and high density lipoprotein cholesterol

^eAge, systolic blood pressure, smoking, diabetes, total and high density lipoprotein cholesterol, c-reactive protein, family history of premature myocardial infarction

Repeating our analysis with only the directly genotyped SNPs had no appreciable effect on our results, nor did excluding the SNPs associated only with C-reactive protein, hemoglobin A1c or triglycerides.

Comment

In this analysis, we constructed two literature based genetic risk scores for cardiovascular disease and tested their relationship to incident cardiovascular events and their potential to improve prediction in a prospective cohort of 19,313 initially healthy white women from the Women’s Genome Health Study. While both the score based on genetic markers for both cardiovascular disease and intermediate phenotypes (101 SNP score) and the score based only on genetic markers for cardiovascular disease (12 SNP score) were associated with increased risk after adjustment for age, the ability of either score alone to discriminate between women at risk for cardiovascular events and those not at risk was minimal with a c-index of 0.52 for both scores. Furthermore, neither genetic risk score remained associated with incident cardiovascular disease after adjustment for traditional risk factors nor had any significant impact on discrimination or reclassification. In contrast, self-reported family history remained associated with incident cardiovascular disease after adjustment for other risk factors and had a substantive effect on reclassification fit.

Previous studies using genetic risk scores for cardiovascular disease have found some evidence of increased prediction.^{5, 6} However, these studies have used genetic markers which replicated in the same population used to test the score rather than a strictly literature based approach, a method that runs the risk of overfitting and consequently yielding overly optimistic results. To avoid this potential bias, we chose to use all genes reported in the literature to be associated with cardiovascular disease or an intermediate phenotype with genome-wide significance. To the extent that the published associations identify useful genetic risk factors, our approach may more accurately reflect the potential of current genetic markers to improve risk prediction on a population basis.

We believe these data to have clinical relevance for several reasons. First, genome-wide testing is currently available and marketed to the general public. Our study finds no clinical utility in a multi-locus panel of SNPs for cardiovascular risk based on the best available literature. Second, our data confirm the utility of intermediate phenotypes such as total cholesterol, high density lipoprotein cholesterol, and blood pressure, in as much as the genetic risk scores were no longer significant after adjustment. This utility most likely reflects the integration of both genetic and environmental factors into measured biomarker levels and to cardiovascular outcomes. Third, our findings confirm the importance of family history of cardiovascular disease, which also integrates not only shared genetics, but also shared behaviors and environmental factors. At the same time, we believe that our data suggest areas for further biomarker research which may improve prediction. To date, almost all genome-wide association studies have been limited to common SNPs, leaving the contribution of rare variants to cardiovascular risk prediction untested. Furthermore, given the continued utility of intermediate phenotypes, the ongoing explorations in metabolomics and proteomics could add significantly to our ability to predict risk.

Limitations of our study merit consideration. As suggested by the strong effect of family history on CVD risk, there is a substantial risk component due to genes and shared environment which may be elucidated by future genetic research. While the NHGRI catalog is based on all available published genome-wide studies, these have focused to date only on common SNPs and thus we were also unable to assess the potential contributions of rare alleles. However, if only discovered through a major increase in sample size, it is possible that these as yet unidentified variants will also have increasingly small effects²⁶. It may also be possible in the future to obtain stable estimates of the exact effect or hazard ratio for use in a weighted score and to find interactions between genes or with genes and other markers, both of which may improve predictive ability.

In conclusion, in this large-scale prospective cohort of white women, a comprehensive literature based genetic risk score - though associated with cardiovascular events after adjustment for age - did not improve cardiovascular risk prediction. This was true whether the component genetic effects were extended to include polymorphisms acting on intermediate phenotypes or restricted only to those directly associated with cardiovascular disease outcomes. While the importance of genetic data in understanding biology and etiology is unchallenged, we did not find evidence in this study of more than 19,000 women to incorporate the current body of known genetic markers into formal clinical tools for cardiovascular risk assessment.