A Genetic Risk Score is Associated with Incident Cardiovascular Disease and Coronary Artery Calcium - The Framingham Heart Study

Abstract

Background

Limited data exist regarding the use of a genetic risk score for predicting risk of incident cardiovascular disease (CVD) in US based samples.

Methods and Results

Using findings from recent GWAS, we constructed genetic risk scores (GRS) comprised of 13 genetic variants associated with myocardial infarction (MI) or other manifestations of CHD and 102 genetic variants associated with CHD or its major risk factors. We also updated the 13 SNP GRS with 16 SNPs recently discovered by GWAS. We estimated the association, discrimination and risk reclassification of each GRS for incident cardiovascular events and for prevalent coronary artery calcium (CAC).

In analyses adjusted for age, sex, CVD risk factors and parental history of CVD, the 13 SNP GRS was significantly associated with incident hard CHD (HR 1.07, 95% CI 1.00-1.15, p=0.04), CVD (hazard ratio [HR] per-allele 1.05, 95% confidence interval [CI] 1.01-1.09; p=0.03), and high CAC (defined as >75^th age and sex-specific percentile; odds ratio [OR] per-allele 1.18, 95% CI 1.11-1.26, p=3.4 × 10^-7). The GRS did not improve discrimination for incident CHD or CVD but led to modest improvements in risk reclassification. However, significant improvements in discrimination and risk reclassification were observed for the prediction of high CAC. The addition of 16 newly discovered SNPs to the 13 SNP GRS did not significantly modify these results.

Conclusions

A GRS comprised of 13 SNPs associated with coronary disease is an independent predictor of cardiovascular events and of high CAC, modestly improves risk reclassification for incident CHD and significant improves discrimination for high CAC. The addition of recently discovered SNPs did not significantly improve the performance of this GRS.

Introduction

Current cardiovascular disease (CVD) risk prediction models based on conventional risk factors perform well and are frequently used to guide treatment decisions. However, nearly 15% of individuals who develop CVD have few risk factors and are deemed to be low risk based on such models¹. In addition, identifying higher risk patients from the large number that are deemed intermediate risk, could also help target individuals for preventative treatment. Efforts to improve CVD risk prediction are needed, given that the first manifestation may be sudden-death and CVD is preventable. Although the familial nature of CVD has been documented for many years^2-4 and the addition of family history has been shown to improve risk prediction^5,6, the genetic variants responsible for the increased familial risk were, until recently, unknown. Genome-wide association studies (GWAS) have uncovered several common genetic variants (single nucleotide polymorphisms, or SNPs) that are robustly associated with myocardial infarction, CHD or CVD risk factors,including dyslipidemia and hypertension, and have been replicated in multiple independent samples⁷. The identification of these genetic variants provides an opportunity to evaluate whether addition of a genetic risk score (GRS) to risk models may improve predictive performance or lead to meaningful changes in risk classification. Recent studies evaluating the utility of adding genetic variants identified in GWAS to cardiovascular risk prediction have provided conflicting results as to the utility of genetic information for CVD risk prediction^8-10. To date, these studies have focused on SNPs reported in GWAS prior to 2011 and have been limited to female health professionals⁸ or to individuals from Scandinavian countries that may not be entirely generalizable to community-dwelling men and women in the US¹⁰. Despite the increase in direct-to-consumer testing for genetic variants for CVD and other chronic diseases in the US¹¹, there are limited data evaluating the utility of adding genetic variants to CVD risk prediction in a U.S. community-based sample of both men and women, where such information would ultimately be used.

Accordingly, we sought to evaluate whether a GRS composed of genetic variants from recent GWAS of MI and CHD that have been reported in other GRS studies, as well as variants identified in GWAS of cardiovascular disease risk factors, is associated with incident CHD, incident CVD and high coronary artery calcium (CAC), a measure of subclinical atherosclerosis^12-15, in Framingham Heart Study participants. We specifically sought to evaluate whether the addition of a GRS improved discrimination or risk reclassification in this community-based sample.

Methods

Study Sample

In 1948, the Framingham Heart Study enrolled 5209 individuals into a longitudinal cohort study. Original cohort participants were examined approximately every two years. Subsequently, in 1971, the Framingham Offspring study enrolled 5124 children and spouse of the children of the original cohort. In 2002, the Framingham Third Generation study enrolled 4,095 children of the Offspring cohort. The study design and entry criteria for each cohort has been described previously^16-18. Participants of the Framingham Offspring were evaluated approximately every 4 years. The sample of Offspring participants used in these analyses has been previously described (baseline examinations 1 and 3, N= 4073 individuals)¹⁹ as well as additional data from the 6^th Offspring examination (N= 301 individuals). Of these individuals, 847 were excluded for lack of DNA (refused to participate (n=795) or poor quality of DNA (n=52)) and an additional 488 participants were excluded for incomplete genotype data or other covariates. The final study sample size that was analyzed for incident events was 3,014 participants contributing 6924 person-exam cycles. Of these participants, 1,164 also participated in a Multi-Detector Computed Tomography (MDCT) sub-study of CAC. Details of the MDCT sub-study have been previously described²⁰.

In a separate analysis of CAC, we analyzed participants from the Third Generation cohort who were also included in the MDCT sub-study. A total of 2,106 Third Generation participants underwent cardiac CT for evaluation of CAC as a measure of subclinical atherosclerosis. Of these, 42 were excluded for prevalent CVD and 102 were excluded for missing covariate data. The total sample analyzed for CAC from the Third Generation MDCT sub-study was 1,962. Although, the Third Generation sample does not contain the same individuals as the Offspring sample described above, individuals from in these 2 samples are genetically related (i.e. individuals in the Third Generation sample are the children of the individuals in the Offspring sample.). This study was approved by the institutional review board at Boston University, and written informed consent was obtained from all participants.

Incident Cardiovascular Disease Events

Based on previously described Framingham Heart Study criteria, the primary outcome was incident general CVD defined as cardiovascular death, MI, coronary insufficiency, angina pectoris, stroke, transient ischemic attack, intermittent claudication, or congestive heart failure during follow-up¹⁹. We also evaluated a more restrictive outcome of hard CHD defined as coronary death or MI and denoted as “hard CHD” in this manuscript. All events were adjudicated by a panel of three physicians, including at least one cardiologist, using pre-specified diagnostic criteria.

Coronary Artery Calcium Measurements

Participants underwent a chest scan on an 8-slice MDCT scanner (LightSpeed Ultra; General Electric, Milwaukee, Wisconsin) for quantification of CAC, as described previously²⁰. Briefly, 48 contiguous 2.5-mm-thick slices were acquired and each participant underwent a second scan after briefly being repositioned on the table. Using a dedicated off-line workstation (Aquarius; Terarecon, San Mateo, California), each image was evaluated for the presence and amount of CAC by an experienced reader. We defined a calcified lesion as an area ≥3 connected pixels with an attenuation >130 Hounsfield units and an Agatston score was calculated as previously described²¹.

Cardiovascular Risk Factors

Each participant underwent a routine medical history, physical examination, and fasting blood draw at each examination. Using conventional methods, total cholesterol and high-density lipoprotein cholesterol (HDL) were determined on fasting blood samples. Cigarette smoking and a complete medication history were obtained by participant interview. Blood pressure was determined in the left arm using a mercury sphygmomanometer in subjects who had been seated for at least five minutes. Adult onset diabetes mellitus was defined as fasting plasma glucose ≥126 mg/dL or treatment with a hypoglycemic agent and/or insulin. Premature parental occurrence of CVD was defined as the occurrence of a validated parental CVD event before age 65 years. At the time of this study, all parents in the original cohort were >65 years of age allowing for complete ascertainment of parental history of CVD in offspring.

Construction of the Genetic Risk Score

We created two genetic risk scores for primary evaluation. The 13 SNP GRS was based on SNPs reported in a recent GWAS of CHD or MI^22-28 and consisted of 13 SNPs that have been robustly associated with this outcome (p<5×10-8) in a discovery sample with subsequent replication in one or more additional independent samples. These SNPs have been included in recent GRS studies of CHD and CVD. Details about these SNPs, listed under the trait “MI”, are provided in the Supplementary Table. Participants were genotyped for 12 of these SNPs on an Illumina platform. The minimum call rate was 99.8%, and all SNPs were in Hardy-Weinberg equilibrium. Strand orientation was determined by comparing alleles and minor allele frequencies with those in the initial discovery report. For the rs3798220 SNP in the LPA locus, we used the genotype from the Affymetrix 500K chip + 50K molecular inversion probe (MIP) chip genotypes available from the SNP Health Resource (SHARe). All SNPs in the final gene score were uncorrelated (r² < 0.3).

We also created a less restrictive GRS for evaluation, based upon available GWAS of major CVD risk factors, that included 102 SNPs (13 CHD/MI SNPs from the 13 SNP GRS, in addition to 89 SNPs associated with the following cardiovascular risk factors (see Supplementary Table 1): low density lipoprotein cholesterol (LDL), high density lipoprotein cholesterol (HDL), triglycerides, diabetes (or fasting plasma glucose), systolic and diastolic blood pressure and c-reactive protein). This score consisted of 85 (82.5%) directly genotyped SNPs. The remaining SNPs were imputed using HapMap data as part of SHARe, as previously described²⁹, and had an imputation quality > 0.65. All 102 SNPs in this final GRS were uncorrelated (r² < 0.3). For any correlated SNPs, a single SNP was selected for a given genomic region by giving preference to genotyped SNPs over imputed SNPs.

Each GRS was generated as a count of the risk alleles for each of the included SNPs (i.e. 2, for homozygous, 1 for heterozygous and 0 for absence of a risk allele) for a total score ranging from 0 to 26 and 0 to 204, for the 13 SNP GRS and 102 SNP GRS, respectively. In secondary analyses we also computed a weighted 13 SNP GRS using the point estimate for the beta coefficient reported in the original report as the weight for each risk SNP and we also conducted a secondary analysis to evaluate a GRS that excluded the LPA SNP. Hazard ratios for incident events using the weighted score or after exclusion of the LPA SNP did not differ substantially from the primary analysis and are therefore not presented.

After initial submission of our manuscript, 17 novel, replicated genome-wide significant SNPs for MI and CHD were reported in 2011 from two large-scale GWAS.^30,31 Therefore, in secondary analyses, we have provided an updated version of our 13 SNP GRS incorporating these variants (Supplementary Table 2). We excluded one of these 17 SNPs for not being in Hardy-Weinberg equilibrium (p=2 × 10^-25) and for a suboptimal call rate (90%) leaving a total of 16 new SNPs and present results for this updated 29 SNP GRS (genotypes for 10 of these SNPs were imputed using HapMap data). For these additional analyses, we excluded an additional 127 individuals who did not have complete genotypes for these additional SNPs, reducing our sample to 2887 participants (1331 men and 1556 women).

Statistical Analysis

Kaplan-Meier rates were calculated for hard CHD and CVD at 10 years of follow-up. For incident events, we used Cox-proportional hazard models to estimate the association between the genetic risk scores and incident events at 10 years. Follow-up began at the baseline examination and participants were censored at death, loss to follow-up, or at the next baseline examination or 12 years, whichever was earlier. Baseline examinations were considered as examinations 1, 3 and 6 and follow-up windows were pooled. Each participant contributed a mean of 2.3 person-examination cycles. This method of pooling person-examinations has been shown to provide valid estimates of association similar to using time-dependent Cox regression as demonstrated by D'Agostino et al³². Separate models were fit for each outcome (hard CHD and CVD). Three covariate adjustments were used: age and sex (model set 1); age, sex and cardiovascular risk factors including total cholesterol, HDL, presence of diabetes, systolic blood pressure (and anti-hypertensive treatment) and cigarette smoking (model set 2); and age, sex, cardiovascular risk factors and parental history of CVD (model set 3). For analyses using CAC as an outcome, we used logistic regression to estimate the association between the GRS and high CAC. High CAC was defined as Agatston score greater than the age and sex-specific 75^th percentile in a healthy reference population (free of clinically apparent CVD or CVD risk factors). Given that CAC is highly correlated with increasing age and male sex, we utilized an age and sex-specific percentile based outcome measure to estimate whether the GRS improved discrimination for accelerated atherosclerosis for a given age and sex stratum. As previously reported, the 75^th percentile Agatston score cut-offs for the <45, 45-54, 55-64, 65-74 and ≥75 age groups were 0, 28, 177, 582, 736 for men; and 0, 0, 26, 43, 356 for women, respectively³³.

All analyses were also performed accounting for the family structure of the sample. Incident event analyses were performed using a Cox regression with a frailty term clustering on family and CAC logistic regression analyses were performed using generalized estimating equations (GEE) clustering on family. Results adjusting for family structure did not materially change effect estimates or standard errors and are therefore not shown. The hazard ratios and 95% confidence intervals are presented for each model.

To evaluate whether the addition of a GRS improved model performance, we evaluated changes in discrimination as quantified by the c statistic, integrated discrimination improvement (IDI) and continuous net reclassification improvement (contNRI) and in risk reclassification measured by the traditional NRI with 3 risk categories (0-6%, 6-20% and >20%), as previously described^,34. Furthermore, we calculated calibration statistics for models with and without GRS using the Nam and D'Agostino modification of the Hosmer and Lemeshow chi-square statistic³⁵. Only the primary 13 SNP GRS and the updated 29 SNP GRS were evaluated in these analyses, as the 102 SNP GRS was not found to be a significant predictor of incident hard CHD or CVD events. To optimize sample size given the missing data on family history status (only 1560 participants had complete validated parental history of CVD from both parents), we assumed that parental history was missing at random and used multiple imputation based on the clinical covariates and GRS to impute family history. Changes in performance measures with the addition of the GRS were evaluated across sets of nested models that included: age and sex (model set 1); age, sex and cardiovascular risk factors including total cholesterol, HDL, presence of diabetes, systolic blood pressure (and anti-hypertensive treatment) and cigarette smoking (model set 2); and age, sex, cardiovascular risk factors and parental history of CVD (model set 3). For risk reclassification, we calculated the NRI and IDI for the addition of the GRS to a clinical model that included age, sex and CVD risk factors. The following predicted risk cut-offs were used for these analyses: 0-6%, 6-20% and >20%. We reported the NRI as a summary measure as well as separately for events and non-events. We also used a novel continuous NRI that is independent of pre-specified risk cut-offs and maximizes statistical power³⁶. For this novel metric, any increase in predicted risk among those who develop events and any decrease among those who do not correspond to improved reclassification. Confidence intervals for NRI and IDI were computed using bootstrap with 999 replications. All analyses were performed in SAS 9.1 (SAS Institute). A two-tailed p-value < 0.05 was used to indicate statistical significance.

Results

We included 1388 men (3168 person-exam cycles) and 1626 women (3756 person-exam cycles) with a mean age of 49±10years in the study sample. Median follow-up was 11 years. Follow-up rate was >85% at 12 years. A parental history of CVD was estimated to be 26%. Clinical characteristics of the sample for the incident cardiovascular disease event analysis are described in Table 1. During follow-up there were 539 CVD events and 182 hard CHD events. The mean 13 SNP GRS was 12.7±2.1 risk alleles and the mean 102 SNP GRS was 108.2±6.3 risk alleles.

Table 1.

Characteristics of Offspring Participants in the Incident Events Analysis

	All participants
N*	3014
Median follow-up (yrs)	11
Age (yrs)	49 ± 10.9
Men (%)	46
Systolic blood pressure	124 ± 16.7
Total cholesterol	206 ± 38.2
HDL cholesterol	52 ± 15.5
Smoking (%)	26
Diabetes (%)	5
HTN treatment (%)	14
Parental occurrence of CVD (%)	26
Kaplan-Meier Event Rate at 10 years of follow-up (n events)
Hard CHD	2% (182)
CVD	7% (539)

All data are presented as mean ± standard deviation unless otherwise specified

^*Data are based on 3168 person-examination cycles from 1388 men and 3756 person-examination cycles from 1626 women.

CHD – coronary heart disease; CVD – cardiovascular disease; HTN - hypertension

Association of Genetic Risk Scores with Incident Cardiovascular Disease Events

After adjustment for CVD risk factors and parental history of CVD, the 13 SNP GRS composed of risk variants for MI/CHD was associated with incident hard CHD (HR 1.07, per risk allele, 95% CI 1.00-1.15; p=0.04) and incident CVD (HR 1.05 per risk allele, 95% CI 1.0- 1.09; p=0.03) events (Table 2). A 103 SNP GRS comprised of the 13 coronary disease variants and an additional 89 variants associated with several cardiac risk factors was not associated with incident CVD or hard CHD events (HR 1.01, 95% CI 0.99-1.03, p=0.48; HR 1.01 per risk allele, 95% CI 1.01-1.02, p=0.54, for hard CHD and CVD, respectively).

Table 2.

Hazard Ratios for the Association between a GRS (per Allele) and Incident Events

	No. of SNPs	Age and sex adjusted HR (95% CI)	P	Incident Hard CHD Risk factor adjusted HR† (95% CI)	P	Fully adjusted HR* (95% CI)	P
GRS 1	13	1.09 (1.02-1.17)	0.02	1.07 (1.00-1.15)	0.04	1.07 (1.00-1.15)	0.04
GRS 2	103	1.02 (0.99-1.04	0.14	1.01 (0.99-1.03)	0.48	1.01 (0.99-1.03)	0.47

	No. of SNPs	Age and sex adjusted HR (95% CI)	P	Incident CVD Risk factor adjusted† (95% CI)	P	Fully adjusted HR* (95% CI)	P
GRS 1	13	1.06 (1.02-1.10)	0.006	1.05 (1.01-1.09)	0.03	1.05 (1.01-1.09)	0.03
GRS 2	103	1.01 (1.00-1.03)	0.07	1.01 (1.00-1.02)	0.54	1.00 (0.99-1.02)	0.52

^†Adjusted for age, sex, total cholesterol, HDL, presence of diabetes, systolic blood pressure (and anti-hypertensive treatment) and cigarette smoking.

^*Adjusted for age, sex, total cholesterol, HDL, presence of diabetes, systolic blood pressure (and anti-hypertensive treatment), cigarette smoking and parental history of CVD.

GRS – genetic risk score; SNPs – single nucleotide polymorphism, HR – hazard ratio; CHD – coronary heart disease; CVD – cardiovascular disease

Changes in Discrimination, Calibration and Risk Reclassification with the Addition of a 13 CHD/MI SNP GRS for the prediction of hard CHD and CVD events

Improvements in discrimination were assessed for the addition of the 13 SNP CHD/MI GRS to three separate models predicting risk of hard CHD and CVD (Table 3). First, we examined whether knowledge of a GRS with limited information, improved the c-statistic. The addition of the GRS to a model including only age and sex (c-statistic = 0.732) offered marginal improvement (c-statistic with GRS = 0.737) for incident CHD (Table 3, Model 1). We also constructed a model predicting hard CHD that included age, sex and cardiovascular risk factors. In this model, the c-statistic increased marginally from 0.819 to 0.822 with the addition of the GRS (Table 3, Model 2). Similar changes in the c-statistic were seen when the GRS was added to a model including the above risk factors and parental history of CVD (change in c-statistic = 0.002; Table 3, Model 3). Results did not materially change when the outcome was incident CVD (Table 3). All models with or without the CHD/MI GRS were well calibrated (data not shown)

Table 3.

C-statistics for the Addition of a Coronary Disease Genetic Risk Score to Models Predicting 10-yr Risk of Cardiovascular Disease or Coronary Heart Disease

	CVD at 10 years		Hard CHD at 10 years
	c	95% CI	C	95% CI
Model 1
age, sex	0.728	0.706-0.750	0.732	0.696-0.768
+GRS	0.730	0.708-0.751	0.737	0.701-0.773
Model 2
age, sex, cigarette smoking, total cholesterol, HDL, systolic blood pressure (and treatment), diabetes	0.786	0.768-0.805	0.819	0.791-0.847
+GRS	0.786	0.768-0.805	0.822	0.794-0.850
Model 3
age, sex, cigarette smoking, total cholesterol, HDL, systolic blood pressure (and treatment), diabetes, parental history of CVD	0.787	0.769-0.806	0.820	0.792-0.848
+GRS	0.788	0.769-0.806	0.822	0.795-0.851

CVD – Cardiovascular disease; CHD – coronary heart disease; GRS – genetic risk score; HDL – high density lipoprotein.

In an assessment of risk reclassification, the addition of the CHD/MI GRS to a model including age, sex and CVD risk factors did not lead to statistically significant changes in the NRI for incident hard CHD or CVD at 10 years using standard risk categories (Table 4). Because increase in the c statistic may not be sensitive as a measure of improvement in model performance and the NRI is dependent on the risk categories selected, we also evaluated the IDI and a novel continuous NRI metric which is independent of risk categories and maximizes statistical power. Based on the IDI, the improvement in separation between events and non-events was minimal for all models considered (Table 4). Using the continuous NRI metric, the addition of a GRS to a model predicting 10-year risk of hard CHD, including age, sex and cardiovascular risk factors led to statistically significant but modest improvement in risk reclassification (NRI = 0.17, 95% CI 0.01-0.33) and these results were not affected when parental history of CVD was included in the baseline model (NRI = 0.19, 95% CI 0.02-0.34).

Table 4.

Risk Reclassification for the Addition of the Coronary Disease Genetic Risk Score to a Model Predicting 10-year Risk of Hard CHD

	NRI		Event/Nonevent NRI	IDI		contNRI
Model 1		95% CI			95% CI		95% CI
age+sex (reference model)	-	-	-	-	-	-	-
+GRS	0.043	(-0.003, 0.088)	0.041/0.002	0.001	(0.001, 0.002)	0.22	(0.057, 0.377)
Model 2
age+sex+ CVD risk factors (reference model)	-	-	-	-	-	-	-
+GRS	0.001	(-0.040, 0.039)	0.0003/0.0005	0.001	(-0.001, 0.003)	0.17	(0.010, 0.328)
Model 3
age+sex+CVD risk factors+parental history (reference model)	-	-	-	-	-	-	-
+GRS	-0.01	(-0.052, 0.033)	-0.011/0.001	0.001	(-0.001, 0.003)	0.19	(0.024, 0.344)

NRI is calculated for the addition of the GRS to a reference model with the following risk cut-offs for 10-year risk of hard CHD: low (<6%), intermediate (6-20%) and high (>20%).

CVD – Cardiovascular disease; GRS – Genetic risk score; NRI – Net reclassification index; IDI – Integarted discrimination index; contNRI denotes a continuous cut-off independent form of the net reclassification index³⁶

Association Between a GRS and CAC and Improvements in Discrimination for High CAC

We examined the association of the 13 SNP CHD/MI GRS with high CAC in the 1,164 Offspring participants who were included in the incident analysis and also had an MDCT scan for assessment of CAC. In fully adjusted models, the odds ratio per allele of GRS for the presence of high CAC was 1.18 (95% CI 1.11-1.26; p=3.4 × 10^-7). There was a two-fold increase in the odds of high CAC for individuals in the highest GRS tertile compared to the lowest (OR 2.04, 95% CI 1.48-2.83) (Figure 1A). Moreover, the addition of the GRS to a model including age, sex, CVD risk factors led to changes in the c-statistic (change in c-statistic = 0.03; from 0.64 to 0.67). The observed improvement in discrimination was not affected when parental history of CVD was included in the baseline model. The addition of the GRS to a model including age, sex and cardiac risk factors led to improvements in both the IDI (p = 1.1 × 10^-7) and the continuous NRI (NRI 0.29; 95% CI 0.17-1.41, p = 2.7 × 10^-7) for the prediction of high CAC.

Figure 1. Fully Adjusted Odds Ratios for High CAC in Offspring Participants Included in the Incident Events Analysis (A) and in Generation 3 Participants (B).

Several major discoveries in the genetics of cardiovascular disease have been made in recent years. However, whether the addition of genetic information can improve current risk prediction models for cardiovascular disease (e.g. Framingham Risk Score) has not been well established. Using data from over 3,000 individuals from the Framingham Offspring Study, we demonstrate that a genetic risk score made up of 13 genetic variants associated with myocardial infarction is associated with both incident coronary heart disease and cardiovascular disease. Furthermore, we also show that the genetic risk score is also associated with high coronary artery calcium, a subclinical marker of atherosclerosis. However, the addition of these genetic variants only led to very modest improvements to risk assessment when added to the Framingham Risk Score. Our results suggest that the addition of currently available genetic information does not appreciably improve the assessment of future cardiovascular risk at the current time. As additional genetic variants are discovered, future iterations of a genetic risk score will require further investigation.

In a separate analysis of 1,962 participants of the Generation 3 cohort who underwent coronary artery calcium scoring, we also observed a marked increase in the prevalence of high CAC across tertiles of the 13 SNP GRS (OR 1.41, 95% CI 1.06-1.89 for highest vs. lowest GRS tertile) which persisted after adjustment for cardiovascular risk factors (Figure 1B). The addition of the 13 SNP GRS to a model including age, sex and cardiac risk factors for prediction of high CAC in Generation 3 participants led to improvement in the c-statistic (change in c- statistic = 0.05, from 0.66 to 0.71), the IDI (p=0.001) and the NRI (0.13; 95% CI 0.03-0.24, p=0.01).

In analyses using the 103 SNP GRS, the GRS remained modestly associated with high CAC in fully adjusted models (OR per-allele 1.03, 95% CI 1.01-1.05; p=0.003) but we noted only marginal improvement in discrimination for high CAC (change in c-statistic = 0.01; from 0.64 to 0.65).

Impact of adding 16 recently discovered SNPs to the 13 SNP GRS for prediction of CVD, CHD and CAC

To evaluate the impact of adding 16 recently discovered SNP to the GRS, we created a new 29 SNP GRS. The mean score was 24.2±3.2 risk alleles. After adjustment for age, sex, cardiovascular risk factors and parental history, the new 29 SNP GRS was associated with CHD (HR per-allele 1.06; 95% CI 1.01-1.11; p=0.02) but was not associated with CVD (HR per-allele 1.00; 0.98-1.03; p=0.86). However, the 29 SNP GRS remained significantly associated with CAC in Offspring and Generation 3 samples (OR per-allele 1.07, 95% CI1.03-1.12; p=7.5 × 10^-4 and 1.04, 1.01-1.08; p=0.01, respectively) after adjustment for age, sex and cardiovascular risk factors.

With respect to discrimination, the addition of the 16 novel SNPs to the original model including age, sex, cardiovascular risk factors and 13 SNP GRS, did not lead to any significant improvement for CVD (no change in c-statistic). For CHD, we noted a marginal improvement in c-statistic (change in c-statistic = 0.001). The addition of parental history to the baseline models did not affect these results. Lastly, for prediction of high CAC, the addition of the 16 SNPs to a baseline model including age, sex, cardiovascular risk factors and the 13 SNP GRS did not significantly improve the c-statistic.

For risk reclassification, results for reclassification metrics using the 29 SNP GRS (data not shown) were similar to the results for the 13 SNP score data presented in Table 4.

Discussion

In this study of Framingham Heart Study participants, a GRS composed of 13 SNPs from recent GWAS of MI and other CHD is significantly associated with incident hard CHD and incident CVD even after adjustment for traditional cardiovascular risk factors and parental history of CVD. We found that for each risk allele there was a 5% increase in the risk of incident hard CHD or incident CVD at 10 years. In contrast, a more complex 103 SNP GRS including SNPs associated with CVD and CVD risk factors was not significantly associated with either incident hard CHD or incident CVD suggesting that the addition of genetic variants for cardiovascular risk factors already captured in risk prediction models may be of limited utility. We also report that the 13 SNP GRS was significantly associated with increased CAC, an important subclinical marker of coronary artery atherosclerosis, which is also an important predictor of future cardiovascular events^12-15. This represents a novel finding with potentially important mechanistic implications for the role of these genetic variants in CVD.

Despite these significant associations, the addition of the GRS to standard risk models did not lead to any significant improvements in discrimination or risk reclassification for incident CVD but did lead to a modest, significant improvement in one index of risk reclassification for incident hard CHD. On the contrary, the addition of a GRS to models for prediction of high CAC did lead to significant improvements in discrimination suggesting that the GRS appears to be predictive for the presence of accelerated atherosclerosis.

We also updated our GRS with 16 novel SNPs recently discovered in GWAS of MI or CHD. The addition of these 16 SNPs led to very marginal improvements in prediction, discrimination or risk reclassification for cardiovascular disease. The absence of improvement in prediction with the larger SNP score may be due to the very low effect sizes of these additional SNPs. Alternatively, these SNPs may not be as specific for the more general CVD and CHD outcomes (which include other vascular events), as compared with the more restrictive MI outcome for which they were discovered. In addition, not all the SNPs were found to be associated individually with hard CHD or CVD (results not shown) in our sample. Future studies will need to address whether future iterations of a GRS incorporating additional genetic risk variants, including both common and low frequency variants, would lead to clinically useful improvements in clinically relevant metrics of risk prediction and whether such information will be most useful for prediction in children, young adults or older individuals.

Several studies have previously incorporated genetic information for the purpose of risk prediction of CVD^{8-10,26,37-45}. However, a number of these studies relied on genetic variants from the candidate gene era, many of which have failed further replication and may represent largely spurious false positive genetic associations limiting the validity of many prior GRS^46,47. Only a few studies have evaluated the utility of adding GWAS-based genetic variants to risk prediction models. Talmud et al. reported that the addition of 9p21 variants led to significant improvements in reclassification⁴³, however these findings have not been confirmed by others⁴². Kathiresan et al, created a GRS using SNPs strongly associated with lipid levels which was associated with a 15% increase in CHD risk per lipid-associated SNP allele, but did not improve discrimination over and above traditional risk factors and showed only modest improvement in risk reclassification⁴⁰. Paynter et al. created two separate risk scores based on SNPs identified from GWAS studies of cardiovascular disease and cardiovascular risk factors, respectively⁸. After adjustment for risk factors, neither GRS was associated with incident CVD in the Women's Genome Health Study, and therefore, further improvements in discrimination and risk reclassification were also not observed. Although our GRS consisted of a similar set of SNPs to that in the study by Paynter et al., a number of important differences exist between these studies that may explain the disparate results. First, the GRS used in our study was composed primarily of in silico genotypes, not imputed genotypes, and we also included variants in LPA that were not included in previous studies. In addition, our community-based population of Framingham Offspring included men and women that may have been at a higher baseline risk for CVD than the relatively young women in the Women's Health Study. Our findings are in agreement with a recent study evaluating the utility of a GRS consisting of a similar set of SNPs in a large community cohort of men and women from Finland and Sweden which demonstrated a significant increase in risk of CVD in individuals in the top GRS quintile but did not improve measures of discrimination or reclassification¹⁰. Although we found similar associations between the GRS and incident events, the modest improvements we report for discrimination and reclassification confirm that genetic testing for cardiovascular risk prediction is currently of limited clinical utility.

To date, limited functional information exists regarding the novel genetic variants associated with coronary artery disease identified from GWAS studies. Animal studies have recently increased our understanding of the 9p21 variant by demonstrating that a knockout of the mouse ortholog of the CDKN2A/2B locus led to a vascular phenotype of accelerated smooth muscle proliferation⁴⁸. Our findings of marked associations with the GRS and high CAC are therefore noteworthy, suggesting that the presence of multiple risk alleles is highly predictive of an increased propensity to coronary atherosclerosis and vascular wall calcification and implying that several risk variants may act via a vascular phenotype captured by CAC.

Thus far, GRS for the prediction of diabetes⁴⁹, breast cancer⁵⁰ and cardiovascular disease^8,10 based on recent GWAS discoveries have not yielded important improvements in risk prediction. Although this has been seen by some as a failure of the concept of personalized medicine using genomics, the relatively low levels of genetic variance explained by the variants discovered to date suggest that much of the genetic predisposition underlying complex disease remains to be discovered⁵¹. Simulation studies have shown that for CVD, 100-400 common SNPs with modest effect sizes would be needed to improve risk prediction^52,53. Fewer SNPs would be needed if a few rare variants with relatively strong effects were also identified. It is therefore encouraging that the limited number of genetic variants identified to date appear strongly associated with clinical events and intermediate phenotypes, such as CAC, and lead to modest improvement in reclassification (for cardiovascular events, using the continuous NRI metric³⁶) and discrimination (for high CAC) when added to models that include traditional cardiovascular risk factors. Our results demonstrate that although major advances have been made in our understanding of the genetics of coronary artery disease and atherosclerosis, genetic screening for CVD is unlikely to be clinically useful at the current time. However, GWAS with larger sample sizes⁵⁴ and other novel approaches including whole-exome⁵⁵ and whole-genome sequencing are currently ongoing and will undoubtedly identify additional variants which may improve future iterations of a GRS leading to clinically useful improvements in risk prediction. In addition, a GRS may also be most useful earlier in the life course, when knowledge of other risk factors is limited⁴⁹. Whether the improvements in prediction at this early age could lead to changes in lifestyle, risk factor development and outcomes by more aggressive primary prevention strategies remains to be seen. Further studies evaluating the utility of genetic information to predict lifetime risk and the impact of providing such estimates to individuals (including cost-benefit analyses) are required.

Our study has a number of strengths including the use of community-based sample of the US population with a long follow-up and the use of primarily in silico genotyped SNPs. However, a number of limitations deserve mention. First, the genotype score relied only on the few SNPs discovered from GWAS which likely account for a small proportion of the genetic variance of CVD. Furthermore, the gene score used for most analyses was constructed using a simple allele count instead of a more complex weighted score based on effect sizes. Although, we also computed a weighted score based on effect sizes from published reports, this score was not significantly better than the allele count score suggesting that the effect sizes may not be of adequate precision for the purpose of risk prediction. Second, we did not consider gene-gene and gene-environment interactions in the prediction models as no specific interaction has been robustly identified in large populations. The addition of interactions rigorously identified in large-scale studies could significantly improve genetic risk prediction. Third, although we had over 500 events for CVD and over 180 events for CHD, due to the relatively low effect sizes for each allele in the GRS, it is possible that the limited impact of the GRS could be due to the relatively low power of our study to detect any improvements with the GRS. Further studies evaluating the clinical utility of adding a GRS in very large samples of individuals are warranted. Fourth, our sample consisted of predominantly white individuals of European descent which limited population stratification; however, the results of the GRS may not apply to other races or ethnicities with different risks for CVD or different allele frequencies.

In summary, we have shown that a GRS composed of 13 SNPs for coronary disease identified in GWAS studies is significantly associated with incident cardiovascular events but provides only marginal incremental information to risk prediction when added to standard cardiovascular risk factors. Nonetheless, the GRS is strongly associated with prevalent high CAC, a marker of subclinical atherosclerosis, and significantly improves the discrimination of individuals for having high CAC even after consideration of standard risk factors. In addition, we also show that the addition of 16 new SNPs recently discovered in GWAS of MI and CHD, do not appreciably improve the performance of a GRS in prediction, discrimination or risk reclassification. Our results suggest that although genetic information has limited clinical utility at the current time for the prediction of future events, the variants identified appear to provide incremental information for the presence of accelerated coronary atherosclerosis and calcification. As additional variants explaining a larger proportion of the genetic variants are identified, future iterations of a GRS will warrant further investigation.