Joint Effects of Common Genetic Variants from Multiple Genes and Pathways On the Risk of Premature Coronary Artery Disease

Abstract

Objective

To discover common variants in 6 lipid metabolic genes and construct and validate a genetic risk score (GRS) based on the joint effects of genetic variants in multiple genes from lipid and other pathobiologic pathways.

Background

Explaining the genetic basis of coronary artery disease (CAD) is incomplete. Discovery and aggregation of genetic variants from multiple pathways may advance this objective.

Methods

Premature CAD cases (N=1,918) and CAD-free controls (N=1,032) were selected from our angiographic registry. In a discovery phase, single nucleotide polymorphisms (SNPs) at 56 loci from internal discovery and external reports were tested for associations with biomarkers and CAD: 28 promising SNPs were then tested jointly for CAD associations, and a genetic risk score (GRS) consisting of SNPs contributing independently was constructed and validated in a replication set of familial cases and population-based controls (N=1,320).

Results

Five variants contributed jointly to CAD prediction in a multigenic GRS model: odds ratio (OR) =1.24 (95% CI 1.16–1.33) per risk allele, p=8.2×10⁻¹¹, adjusted OR=2.03 (1.53–2.70), 4th vs. 1^st quartile. GRS₅ had minor impact on area under the receiver-operating characteristic curve (p>0.05) but resulted in substantial net reclassification improvement: 0.16 overall, 0.28 in intermediate risk patients (both p<0.0001). GRS₅ predicted familial CAD with similar magnitude in the validation set.

Conclusions

CorGen demonstrates the ability of a multigenic, multipathway GRS to improve discrimination of angiographic CAD. Genetic risk scores promise to increase understanding of the genetic basis of CAD and improve identification of individuals at increased CAD risk.

INTRODUCTION

Much of the genetic basis of coronary heart disease (CHD) remains to be discovered ¹. However, steady progress is occurring using both candidate gene and genome-wide association studies (GWAS)^2–3. Insights into pathobiologic pathways has guided candidate gene testing^{2, 4}, whereas GWAS makes no a priori assumptions about genetic site and provides broader, genome-wide coverage ³. We hypothesized that combining discoveries from both of these complementary approaches and extending consideration to multiple pathobiologic pathways could further advance this effort.

To date, the risk attributable to any individual variant has been modest. However, discovering and combining multiple loci with modest effects into a global genetic risk score (GRS) could improve the identification of high-risk populations and improve individual risk assessment ^5–7. Only a limited number of GRS studies for CHD have been reported, and these have focused mostly on lipid-related genes and especially on low density lipoprotein cholesterol (LDL-C) related variants ^{5, 7}. However, CHD pathogenesis involves multiple stages and other mechanisms and biopathways, including HDL-C, inflammation, thrombosis, and vascular development ^{1–2, 4}.

METHODS

Study Objectives

The primary objectives of the Intermountain Healthcare’s Coronary Genetics (CorGen) study were: 1) to discover all common SNPs among a set of 6 key genes in the reverse cholesterol transport system and test them for associations with angiographic CAD, and 2) to construct and validate a multivariant GRS, based on the joint effects of these and other (literature reported) genes in lipid and 3 other pathobiologic pathways, to discriminate CAD using precise angiographic phenotyping.

Study Participants

Study subjects for the primary association study were selected from Intermountain Healthcare’s ongoing Angiographic Registry and DNA Bank⁸. This Registry is approved by the hospital’s institutional review board, and study subjects give written consent prior to enrollment. Fasting blood is sampled at the time of angiography; DNA and plasma are extracted and stored; demographic and angiographic information is collected and entered into an electronic database; and patients are followed prospectively^8–9.

To optimize genetic susceptibility, we studied younger subjects: men aged ≤60, women ≤70 years. Approximately 3000 subjects (~2000 CAD cases and ~1000 angiographically normal controls, matched 2:1 for sex, age, and date of registry entry) were selected. Clinical diagnoses preceding angiography included stable disease (angina equivalent or other) in 56%, unstable angina in 25%, and acute MI in 19%.

A separate set of cases with highly familial premature CAD (first degree relative with CHD onset <55 in men, <65 in women) from the University of Utah Cardiovascular Genetics Family Tree Registry¹ and a separate set of controls (randomly invited from a public records database) were enrolled as a replication set¹⁰.

Study Design and Selection of Genetic Variants

The study consisted of 3 discovery phases and a validation phase (Supplementary Figure 1), each consisting of separate, independent datasets. In discovery phase 1 (SNP discovery), 6 genes with key roles in reverse cholesterol transport, (cholesteryl ester transfer protein [CETP], hepatic lipase [LIPC], lipoprotein lipase [LPL], lecithin-cholesterol acyltransferase [LCAT], scavenger receptor class B type I [SR-BI], and apo-lipoprotein F [apo F]) were scanned in 62 volunteers; 81 SNPs were discovered, and all variant genomic segments were sequenced. In discovery phase 2 (haplotype or tagging [t]SNP discovery), a separate set of 339 Euro-American volunteers was tested to establish tSNPs using the Horne and Camp method of principal components analysis (supplemental materials and ^11–12). A total of 38 tSNPs were determined to characterize the linkage disequilibrium groups of the 6 genes. Of these, 10 found to be associated univariably with lipoprotein markers or nominally (p<0.2) with CAD in a discovery set of cases (n=915) and controls (n=522) (discovery phase 3) were selected for multivariable GRS modeling.

To complement and expand GRS candidates to other pathways and genome-wide, literature studies were reviewed as of April 2008. Eighteen SNPs showing the most robust and independent associations at genome-wide significance (p<5 × 10⁻⁸) with a CAD-related biomarker (i.e., lipids—divided among LDL-C, HDL-C, TG—or CRP or MCP1) and/or CAD per se (i.e., 9p21.3) and diversified among lipid/lipoprotein, thrombosis, inflammation, vascular function/unassigned pathways were added to the 10 internal candidates SNPs for GRS modeling (Supplementary Figure 1, Supplemental Table 1).

The risk allele for SNP candidates was designated a priori based on associations with a recognized risk marker (e.g., lipid levels) or, preferentially, when reported (for literature SNPs) or angiographically determined (for internal SNP candidates), on associations with CAD directly, prior to entry into multivariable genetic risk score modeling.

Multivariable GRS modeling using backward logistic regression then was used in an angiographically-phenotyped set of 1,918 cases and 1,032 controls to reduce the 28 candidate SNPs from the 2 sources to a set of SNPs contributing jointly to CAD prediction.

In validation testing, the reduced set GRS determined in the discovery set was prospectively tested in a completely independent set of familial CAD cases (n=312) and population controls (n=1,008).

Definition of Angiographic Characteristics and Clinical Covariables

The presence of CAD was determined by coronary angiographic analysis masked to genetic information. Patients were categorized as free of CAD (no lesions noted angiographically), mild/moderate CAD (i.e., most severe lesion <70% stenosis), or severe CAD (i.e., at least one lesion ≥70% stenosis). Patients with mild/moderate CAD were excluded as indeterminate.

Standard clinical criteria and specific therapies were used to define the presence of diabetes, hypertension, and hyperlipidemia⁹. Smoking was defined as current smoking or a >10 pack-year smoking history. Family history was positive if a parent, sibling, or child manifested CAD or myocardial infarction by age 55 in males or 65 in females.

DNA Extraction and Genotyping

DNA was extracted from blood samples, quality and quantity assessed, and genotyping performed using standard techniques as detailed in Supplementary Materials. Call rates for SNPs in the GRS were 95–97% and required to be >90% for all tested SNPs.

Computation of Genetic Risk Scores

Two methods were used to create the multivariable GRS: a simple, unweighted count method (count GRS) and a weighted method (weighted GRS) ^{6, 13}. Both methods assumed each SNP to be independently associated with risk. (The independence of SNP associations with CAD/CAD risk markers was recently tested in this dataset and found to be valid ¹⁴). An additive genetic model was assumed: weightings of 0, 1, and 2 were given according to the number of risk alleles present¹⁵.

The count method assumed that each SNP contributed equally to CAD risk and was calculated by summing the number of risk alleles across the panel of SNPs tested. This produced a score between 0 and twice the number of SNPs, i.e., representing the total number of risk alleles. The weighted GRS was calculated by multiplying each ϐ-coefficient for the CAD phenotype from the discovery set by the number of corresponding risk alleles (0, 1, or 2) and then summing the products. The GRS was modeled as a continuous variable and as quartiles.

Statistical Analysis

Statistical analyses were performed using SPSS version 15.0 (Chicago, IL). Chi-square tests (Armitage 1 df test-of-trend for additive genetic model) and t tests were used for comparing proportions and means, respectively, between cases and controls. Logistic regression was used to determine the effect of each variant separately and combined on risk for angiographic CAD. Odds ratios (OR) and 95% confidence intervals (CI) are reported for the high- versus low-risk alleles assuming an additive risk model. Multivariable analyses were used to adjust for history of hypertension, hyperlipidemia, smoking, diabetes, family history, ethnicity/race, and body mass index (BMI). Sex and age were matched by design. Power calculations used nQuery Advisor v.4.0 (Statistical Solutions, Saugus, MA). For risk alleles with minor allele frequency ≥10%, the study had 80% power to detect an OR for CAD of ≥1.3 and >90% for OR>=1.35, for 2-sided alpha of <0.05 for 2000 cases and 1000 controls. Associations with CAD for individual SNPs were considered significant at p≤0.05, and in aggregate for GRS models at p≤0.01.

We plotted receiver-operating characteristic (ROC) curves and calculated areas under the curve (AUC) for logistic regression models including conventional risk factors without and with GRS¹⁶. We classified Framingham risk scores (FRS) into 4 categories, with intermediate risk categories defined as 10-year risks of 5 to 9% (low-intermediate) and 10% to 19% (high-intermediate)^17–18. The potential of GRS to improve individual risk stratification then was measured using the net reclassification improvement (NRI) method, using flow-limiting CAD (>70% stenosis) as the clinical CHD equivalent and excluding patients with diabetes (an a priori high-risk equivalent and not included in FRS scoring) ^{16, 19–20}.

Sources of funding

This study was funded by a grant from the National Heart, Lung, and Blood Institute, R01- HL071878. The authors are solely responsible for the design and conduct of this study, all study analyses, and drafting and editing of the paper.

RESULTS

Patient Characteristics

Characteristics of the angiographic set of cases and controls are summarized in Table 1. Age averaged 53 years, and 1/3 were women. By design, cases were matched to controls for age and sex. Other traditional risk factors were more prevalent in cases. However, (treated) lipid profiles and blood pressures were similar in cases and controls.

Table I.

Characteristics of Discovery Set Angiographic CAD Cases and Controls

Variable	Angiographic	Angiographic
	CAD Cases	Normal Controls
N	1,947	1,036
Age (y) (mean [SD])	53.1±8.0	53.2±8.2
Sex (% women)	35%	36%
Race/ethnicity† (% White)	94%	94%
BMI (kg/m2)	30.1±6.2	29.9±6.5
H/o Hyperlipidemia (%)	67%*	34%
H/o Hypertension (%)	61%*	44%
H/o Diabetes (%)	28%*	12%
H/o Smoking (%)	26%*	14%
Family History CHD	50%*	29%
Systolic BP (mmHg)	140±25*	138±22
Diastolic BP (mmHg)	82±14	81±14
Total Cholesterol (mg/dL)	193±52*	185±44
Triglycerides (mg/dL)	209±196*	181±152
LDL-C (mg/dL)	111±40	108±37
HDL-C (mg/dL)	38.0±12.7*	41.6±13.7
Glucose (mg/dL)	131±67*	110±45

^*p<0.05 for cases vs. controls.

^†No blacks included.

SNP Discovery and Association with CAD

A total of 38 tSNPs in the 6 key lipoprotein metabolic genes were identified by exhaustive scanning (Supplementary Figure 1, Supplementary Table 1); of these, 10 were significantly associated with lipid levels (i.e., at p≤0.002) or at least nominally (p<0.2) with CAD. For initial GRS modeling, these 10 were added to 18 leading literature SNPs reported from GWAS to be associated with CAD-related biomarkers or CAD-risk. These 28 SNPs, with allelic frequencies in cases and controls, are shown in Supplementary Table 1. All SNPs were in Hardy-Weinberg equilibrium (i.e., p>0.05 per SNP).

Genetic Risk Score Modeling

The multivariable genetic model could be simplified by operator-interactive stepwise elimination from a 28- to a 5-SNP genetic risk score (GRS₅) without loss of discrimination (Table 2). Five other SNPs showed multivariable association trends (0.05<p<0.2) but were eliminated in the final model (Table 2, footnote). The Hosmer-Lemeshow statistics for the 5 SNP model (7.92, p=0.24, 6 df) and for the model using GRS₅ as a single variable (3.91, p=0.27, 3 d.f.) both suggested a good fit to the data.

Table II.

Logistic Regression Model for Genetic Risk Score-5 in Angiographic Case/Control Set

Gene / locus (rs)	Risk Allele	ϐ	Sig	OR (95% CI)
CELSR2 / rs599839	Major	0.223	<0.001	1.25 (1.11–1.40)
9p21.3 / rs2383206	Minor	0.191	<0.001	1.21 (1.09–1.34)
CETP / rs289715*	Major	0.217	0.005	1.24 (1.07–1.44)
ApoF / rs78739461*	Minor	0.307	0.015	1.36 (1.06–1.74)
F2 (PT) / rs1799963	Major	0.474	0.046	1.61 (1.01–2.56)

ϐ =beta coefficient; Sig= significance (p-value); OR=odds ratio for angiographic CAD; CI=confidence interval.

^*Internally-discovered SNP. Candidate SNPs contributing trend associations in the multivariable model (0.05>p<0.2) but eliminated in the final model were: ApoC1/E rs4420638, CETP rs180075, CRP rs2794520, GCKR rs780094, and LIPC rs36041167. Effect estimates are for all variables considered simultaneously in the model.

The simple count method GRS₅ yielded an OR=1.24 per risk allele (CI, 1.16–1.33, p=8.2×10⁻¹¹; model χ²_mv=44.8, model χ²_GRS-5 =42.8)(Table 2). GRS₅ was predictive in diabetic (OR 1.34 [CI 1.13–1.59], N=668) as well as non-diabetic subgroups of cases/controls (OR 1.23 [CI 1.14–1.32], N=2282).

Results for continuous GRS₅ and for quartile GRS₅ are shown in Table 3. Angiographic CAD prevalence across the spectrum of count-GRS₅ scores is presented in Figure 1. Comparisons of CAD prevalence in fourth versus first quartile subjects yielded an unadjusted OR=2.07 (CI 1.59, 2.70) and a risk factor adjusted OR=2.03 (1.53–2.70) for count-GRS₅ (Table 3). Weighted GRS₅ results were similar: unadjusted OR=2.06 (CI 1.61–2.65), adjusted OR=2.02 (CI 1.55–2.64).

Table III.

Associations between Count Genetic Risk Score-5 and Angiographic CAD

Count GRS	Continuous GRS	Quartile of Continuous GRS
		1	2	3	4
No. Subjects	2,983	630	969	930	454
Median GRS (Range)	6 (2–10)	5 (2–5)	6 (6)	7 (7)	8 (8–10)
OR (95% CI)/allele	1.24 (1.16, 1.33)	1.0 (-----)	1.12 (0.91, 1.37)	1.54 (1.24, 1.90)	2.07 (1.59, 2.70)
Adjusted*	1.21 (1.13, 1.30)	1.0 (-----)	1.20 (0.96, 1.49)	1.49 (1.18, 1.87)	2.03 (1.53, 2.70)

^*Adjusted for hyperlipidemia, hypertension, BMI, diabetes, smoking, family history. (Cases, controls matched for age, sex by design.)

Figure 1. Prevalence of Angiographic CAD by Count GRS₅ Score.

Percent CAD across GRS₅ categories was highly significant (p-trend=8.2×10⁻¹¹). Reference mean count GRS score is 6.3.

GRS Validation Testing

For validation testing, the GRS₅ was prospectively tested in a completely separate set of 312 unrelated familial coronary disease cases and 1,008 apparently healthy population controls (Supplementary Table 2). In this independent familial case/control set, the GRS₅ remained highly significantly predictive of coronary disease, with a similar magnitude to that observed in the angiographic case/control set, i.e., OR 1.23 (CI 1.11–1.38) per risk allele, p=2.0×10⁻⁴. Also, each risk allele except the rare ApoF variant individually predicted risk in these familial cases, with the common CELSR2 and 9p21.3 polymorphisms showing the strongest associations (Supplementary Table 3).

In addition, an internal analysis of reliability within the primary case/control set was performed by dividing the set into equally sized “discovery” and “replication” subsets. This analysis demonstrated a high degree of internal reliability, with GRS₅ showing consistent and highly significant predictive value in both subsets (ORs 1.26, 1.23, respectively, both p<10⁻⁵).

Incremental Value of Multiple Source SNPs in GRS Modeling

Of the 5 SNPs meeting criteria for GRS₅ membership, 2 (i.e., ApoF rs78739461* and CETP rs289715), one novel*, were from the internal lipoprotein gene discovery effort, and 3 were from external candidates from other loci/pathways (i.e., 9p21 rs2383206, CELSR2/PSRC1 rs599839, F2 rs1799963). SNPs from both sources contributed in a complementary fashion to overall risk prediction; both a GRS₃ (limited to the 3 external SNPs; χ²_GRS-3 =30.6) and a GRS₂ (2 internal SNPs; χ²_GRS-5 =13.6) predicted CAD but were inferior to the combined GRS₅ (χ²_GRS-5 =42.8).

Multivariable Predictive Model

A multivariable predictive model for angiographic CAD incorporating genetic and clinical information is presented in Table 4. The contribution of GRS₅ is intermediate, similar to family history and greater than hypertension. (The impact of age and sex cannot be determined, given the matching design, and the standard use of statins in CAD patients likely impacts the diagnosis of hyperlipidemia.)

Table IV.

Multivariable Predictive Model for CAD Using Genetic Variables and Standard Risk Factors

Variable	Wald χ2	OR (CI)	p Value
Hyperlipidemia	123.51	2.75 (2.30, 3.29)	1.1×10−28
Diabetes	46.12	2.21 (1.76, 2.77)	1.1×10−11
Smoking	39.12	1.98 (1.60, 2.46)	4.0×10−10
Family History	30.01	1.64 (1.37, 1.96)	4.3×10−8
GRS5 / risk allele	28.66	1.21 (1.13, 1.30)	8.7×10−8
Hypertension	2.71	1.16 (0.97, 1.39)	0.10
BMI (per kg/m2)	0.02	1.001 (0.985, 1.018)	0.88*

^*BMI was eliminated from the final model. Effect estimates are for all variables considered simultaneously.

Incremental Value of GRS in Individual Risk Assessment

When count GRS₅ was added to Framingham Risk Score in the full angiographic set, net reclassification fraction was found to be 0.233 in CAD cases (p<0.0001) and 0.073 in no-CAD controls (p=0.005), which when combined yielded a net reclassification improvement (NRI) of 0.160 (p<0.0001)(Supplementary Table 3). NRI restricted to intermediate risk subjects yielded an NRI of 0.283 (p<0.0001)(Supplementary Table 4)^17–18.

As with others’ experience with novel (and most traditional) risk factors^{7, 13, 18}, the addition of GRS₅ had only minor impact on area under the receiver-operating characteristic curve: conventional model C-statistic=0.723 (CI 0.703–0.742), GRS-augmented C-statistic=0.731 (CI 0.712–0.750) (p>0.05).

DISCUSSION

Summary of Key Study Results

CorGen demonstrates the feasibility and potential utility of simultaneously considering the joint effects of common genetic variants from multiple pathobiologic pathways, aggregated as a GRS, to predict the risk of premature CAD. CorGen also demonstrates the complementary effects of combining tSNPs derived from high-definition scanning of candidate genes (i.e., those associated with lipoprotein metabolism) with biomarker-and other risk-related SNPs discovered through high-density GWAS.

Features of CorGen that provide assurance of a valid result include the angiographic characterization of CAD (defining a precise phenotype) and replication in an independent set of cases and controls. On average, each high-risk allele increased risk by 24%, and considered jointly as GRS₅, a fourth quartile score increased CAD risk by over 2-fold. GRS also contributed independently to standard risk factors in multivariable modeling. GRS improved the Framingham risk score category in a net of 16% of subjects overall, which compares favorably with the 12% reported NRI for HDL cholesterol^16–17 and 10% NRI for systolic blood pressure²⁰. Further, application restricted to subjects in the 2 intermediate risk categories improved classification in a net of 28%. These findings suggest its potential clinical utility in individual risk classification, despite (as with others’ experience) its minor impact on population-level AUC^{7, 13, 17, 20}.

Literature Comparisons

Approximately one-half of CHD risk appears to be genetically transmitted¹. However, the proportion of variation among markers of human traits and diseases (including CAD) attributed to individual SNPs has been modest ^21–22, leading to the “common disease, common variant hypothesis”²³. Combining these SNPs to form a more powerful risk predictor underlies the GRS concept. GRS modeling for CHD is in its infancy, but a few reports have appeared, mostly modeling plasma lipids^{5, 7, 13}. CorGen extends these models to multiple additional genes and pathways identified by either candidate-gene or GWAS methods and introduces new markers.

Accounting for Genetic Susceptibility

Uncovering the genetic basis of CHD is an unrealized goal^{1, 10}. The pathophysiology of CHD is a complex interaction of genetic and environmental factors acting directly and indirectly on multiple disease stages from preclinical to clinical CAD to acute coronary syndromes, each stage with a distinct set of risk factors¹. Despite methodological advances such as GWAS, progress has been slow, and well-validated associations remain few, with modest impact. This emphasizes the need for broadening the genetic search, more precisely defining the coronary phenotype, and aggregating individual genetic risk markers into an overall metric.

A major success in this effort has been the discovery and replication of a risk locus at chromosome 9p21.3 (Supplementary references). Although its mechanism remains to be precisely defined, it appears to be involved in vascular development and function, increasing susceptibility to CAD rather than precipitating MI^{10, 24–25}. Some but not other studies have suggested that knowledge of 9p21 status may improve individual risk classification^26–28. In CorGen, 9p21 emerged as a major contributor to the multivariant GRS.

Three other contributors to GRS₅ are involved with lipoprotein metabolism, but in contrast to a previous report⁷, these were not restricted to LDL-C–related genes. In addition, a less common variant from the thrombosis pathway (F2) with a relatively large effect was selected. Such uncommon SNPs have been suggested as a focus for future discovery efforts^{12, 21}. A variant representing vascular inflammation (CRP rs2794520) demonstrated a preliminary association but was eliminated in the final model. Future research should aim to discover additional genetic risk contributors, both universal and population-specific, including those from vascular inflammation and thrombosis pathways.

Study Implications

Here we show that a multivariant GRS can discriminate CAD as well as or better than many standard non-genetic tests and may improve individual risk assessment. In contrast, GRS₅ added little to the AUC for the ROC curve, i.e., did not importantly improve risk prediction over the Framingham Risk Score at a population level. This dichotomy also has been reported for other risk predictors^{7, 13, 18, 26}.

We view CorGen as a proof-of-concept study. Very recent²⁹ and future genetic studies may identify SNPs that add to and refine the CorGen GRS. Nevertheless, of the many high-profile literature SNPs associated with lipids, other disease markers, or even CAD already reported that we tested, only 5 of 28 leading candidates contributed independently to CAD prediction in multivariable GRS modeling. Similarly, of 38 tSNPs discovered internally by extensive scanning to characterize variation in the 6 key lipoprotein metabolic genes, only 2 contributed to the final GRS model. Hence, the goal to account for the greater part of the genetic basis of CAD remains challenging.

GRS may be of particular value in younger cohorts in whom traditional risk factors have not developed and who may benefit from closer surveillance and more aggressive preventive measures. GRS, as other novel risk predictors, may be of greatest incremental value in those at intermediate pre-test probability^{18–19, 30}. Finally, by interfacing genetic loci from multiple pathways, additional insights into disease pathogenesis and treatment targets may be expected.

Strengths and Limitations

The study though prospective in enrollment and hypothesis possesses the limitations of observational studies, including the possibility of uncorrected confounding. The predictive value of the GRS is restricted to angiographic CAD and not myocardial infarction per se. This study focused on a younger population of Euro-Americans, minimizing population stratification and heterogeneity but potentially limiting applicability to other racial/ethnic groups (i.e., African-Americans) and older ages. Risk-associated SNPs while valuable for disease discrimination may not represent casual variants. Although moderately large, CorGen has limited power to discern associations with small effect sizes. However, small-effect variants are unlikely to have important clinical impact or be cost effective for clinical application. Although the GRS₅ was validated internally, external validation in geographically distinct populations is needed, as is further testing of the novel and rare apoF variant. Here we test GRS in the context of the commonly used ATP-III version of the Framingham Risk Score; other risk stratification methods (e.g., FRS including diabetics; Diamond Forrester score) may deserve testing. Given our case-control study design, we are limited to discriminatory analysis and technically cannot precisely estimate the predictive power of GRS for CAD risk. Finally, we acknowledge that our highlighting of a pathway-based approach is primarily conceptual; this effort represents an initial, selective rather than a thorough, definitive application of a pathway approach.

Conclusions

Using a multiple step study design, CorGen has validated the ability of a GRS derived from 5 SNPs to predict premature CAD and has demonstrated the complementary nature of candidate-gene and GWAS approaches. The GRS₅ model provided greater discrimination than any single variant, predicted a 24% risk increment per allele, identified a highest quartile GRS with a two-fold increase in CAD risk, and improved net risk classification in 28% of intermediate-risk individuals. Thus, CorGen demonstrates proof-of-concept that genetic risk scores are a feasible and promising approach to account for the genetic basis of CAD and identify individuals at increased CAD risk.