Genetics of Coronary Artery Disease

Coronary artery disease (CAD) has important genetic underpinnings considered equivalent to that of environmental factors. The heritability of CAD has been estimated between 40% and 60%, on the basis of family and twin studies, a method that yields high precision despite potential bias (Vinkhuyzen et al¹). In the Framingham Offspring Study, the age-specific incidence of CAD increased by >2-fold after adjustment for conventional CAD risk factors in participants with a family history of premature disease.² The Swedish Twin registry reported on close to 21 000 subjects followed up for >35 years and calculated the heritability of fatal CAD events to be 0.57 and 0.38, for men and women, respectively. Of note, heritable effects are most manifest in younger individuals.³ This accords with other data, indicating that the genetic influence is the greatest for early-onset CAD events.⁴

Unraveling the genetic basis of CAD has evolved slowly during the past half-century. Candidate genes coding for proteins of known biological significance in a disease process seemed to provide a logical first step in understanding the genetics of common disease states. Populations of affected and unaffected individuals could be studied by genotyping common single-nucleotide polymorphisms (SNPs) within a gene and its regulatory sequences. Although economically attractive, this approach had many limitations. By definition, studies are limited to genes with a known or suspected role in defining a given phenotype and do not provide new insight into biological pathways leading to disease. Furthermore, candidate gene associations generally failed to replicate for multiple reasons, including statistical power related to inadequate, sample size, heterogeneity of causality, and population stratification.⁵ With the exception of rare monogenic disorders of lipid metabolism, such as familial hypercholesterolemia, there is as yet little support for a role of single-gene disorders in coronary atherosclerosis or plaque rupture as discussed below.

Recent progress in understanding the genetics of CAD and other complex diseases has been driven by technological advances, including high-throughput DNA microarray technology using chips containing up to a million DNA markers consisting of single-nucleotide polymorphisms (SNPs). In the commercial arrays used for genome-wide association studies (GWASs), SNPs (generally 0.5–1 mol/L) are used to tag common variation (SNPs present in ≥5% of the population) across the human genome. Importantly, these are tag SNPs that point to a causative locus but are rarely functional variants. This approach makes use of linkage disequilibrium (LD) that is the nonrandom coinheritance of genetic variants across the human genome. Comparison of the allele frequency of each SNP in cases and controls as part of a GWAS provides an agnostic approach that involves no underlying assumption on candidate genes or pathways. However, the statistical bar for association is high; comparison of a million markers implies a P value of ≤5×10⁻⁸ for GWAS significance. Thus, success has been contingent on the allele frequency and effect size of a given variant and the recruitment of thousands of carefully phenotyped individuals with or without clinical evidence of CAD and collaboration among many groups across the world.

The first robust association with CAD identified by the GWAS approach, a 53-kb LD block containing multiple highly correlated SNPs at 9p21.3 locus, was identified by 3 independent groups in 2007.^6–8 The early discovery of this risk locus was facilitated by its large effect size (>1.3) and risk allele frequency (≈0.48). Approximately 25% of Europeans carry 2 copies of the risk allele and have a 40% increased risk of CAD in general and a 2-fold risk of premature CAD. Here, it is notable that in the Ottawa Heart Study,⁶ an extreme phenotype approach was used comparing young patients with multivessel CAD to elderly, physically active, asymptomatic subjects. Although the discovery population consisted of only 322 cases and 311 controls, the lead SNP rs10757274 achieved a P value of 3.7×10⁻⁶. Consistently, in the large Coronary Artery Disease Genome-wide Replication and Meta-Analysis (CARDIoGRAM) study, the allele-specific odds ratio (OR) for CAD in subjects with onset before 50 years of age was 1.41 (confidence interval, 1.34–1.48) versus 1.24 (confidence interval, 1.20–1.28)⁴ for older individuals.⁴ The 9p21 locus is also associated with the extent and severity of atherosclerosis^9,10 and with a higher risk allele frequency in subjects with multivessel disease.⁹ It has been consistently shown that the risk conferred by this locus is independent of known risk factors, including plasma lipids, blood pressure, diabetes mellitus, obesity, markers of inflammation, age, and sex. Several other vascular phenotypes are associated with the 9p21 risk alleles, including carotid atherosclerosis,¹¹ stroke,^12,13 peripheral arterial disease,¹⁴ and abdominal aortic aneurysm¹⁵ consistent with an effect on atherosclerosis. However, there is an important association with intracranial aneurysms,¹⁶ thus highlighting the possible effects on vascular wall integrity. Distinct haplotype blocks are associated with platelet reactivity¹⁷ and myocardial infarction versus atherosclerosis.¹⁸ There is also a surprising association with periodontitis.^19,20

Since the 9p21 discovery in 2007, large meta-analyses of additional GWASs, the majority of which have been conducted in individuals of European descent, have identified additional loci of smaller effect size but with genome-wide (P<5×10⁻⁸) significance. This success has been built on large collaborative efforts, including the Myocardial Infarction Genomics Consortium,²¹ the CARDIoGRAM consortium,⁴ the Coronary Artery Disease (C4D) Genetics Consortium,²² CARDIoGRAMplusC4D,²³ and others. The risk loci identified include genes known to function in lipoprotein metabolism, hypertension, and other CAD-associated phenotypes but, importantly, include several novel regions harboring genes of unknown relevance to atherosclerosis and plaque rupture, thus highlighting the discovery potential of the GWAS approach. Several risk regions exhibit pleiotropic effects; for example, ABO and SH2B3 associate with multiple CAD and non–CAD-related phenotypes. With the exception of the LPA gene²⁴ encoding lipoprotein(a) and the 9p21.3 locus, the allele-specific ORs for CAD of replicated loci are <1.15, as might be expected for common variants affecting a complex trait.

GWASs of CAD have been designed to test the common disease–common variant hypothesis, and most have applied imputation from HapMap training sets interrogating mainly SNPs present in >5% of the population (minor allele frequency [MAF]>0.05). In contrast to HapMap, the 1000 Genomes Project phase 1, version 3 data set includes >38 million variants; half of which are present in <1 in 200 individuals (MAF<0.005), as well as insertions and deletions (indels). These data were leveraged in a recent meta-analysis led by Martin Farrall that included >185 000 individuals (60 801 CAD cases and 123 504 controls) from 48 studies providing imputed data on 9.4 million SNPs, including 2.7 million low-frequency SNPs (MAF, 0.005–0.05). The summary Manhattan plot is shown in Figure 1. By using both a recessive and an additive model of inheritance, 2 novel loci were identified that would have been missed by conventional analysis. In total, this analysis identified 10 novel CAD loci at GWAS significance, bringing the total number of replicated loci in European populations to 58 (Table 1). Here, it is notable that <25% of the significant loci are related to known CAD risk factors, highlighting the discovery potential of the GWAS approach. Joint association analysis using genome-wide complex trait analysis software revealed no evidence that the signal conferred by common SNPs was because of synthetic associations, that is, linkage with a rare variant(s) of high effect size.²⁵

The 48 GWAS significant loci previously identified explained only ≈11% of the estimated heritability of CAD. Here, beyond considering only GWAS significant loci (P≤5×10⁻⁸), an approximate joint association analysis was performed using genome-wide complex trait analysis software^26,27 to identify 202 false discovery rate variants (q<0.05) in 129 loci. Thus, several loci contained multiple independent signals for CAD association. Together, the 202 independent variants explained ≈28% of predicted CAD heritability.²⁵ Of these, 15 were of low frequency (MAF<0.05) and explained only ≈2% of CAD heritability. Here, it is important to acknowledge that GWAS analysis based on SNP array data has limited power to resolve genes with rare mutation burdens.²⁵ Overall, this analysis not only provided new information on the genetic architecture of CAD but also strongly supports the conclusion that genetic susceptibility to CAD is largely derived from the effect of multiple common SNPs of small effect size. The 202 independent variants reaching a false discovery rate of q<0.05 showed independent enrichment across 11 cell types for histone/chromatin modifications, DNase I hypersensitive sites consistent with other data that GWAS signals are enriched in regulatory regions of the genome.^28,29

The majority of GWASs for CAD have been carried out in populations of European white ancestry with smaller but important studies in East Asian, South Asian, and black populations. Haplotype blocks are segments of DNA that are known and shared within ancestral groups. Because of the recent arrival of humans into Europe, individuals of European ancestry are more similar to each other and have more correlated SNPs and longer haplotype blocks when compared with the more ancient and genetically diverse populations of African ancestry.^30–32 The average haplotype block is ≈20.7 kb in European whites, ≈8.8 kb in blacks, and ≈ 25.2 kb in Han Chinese. Thus, fewer tag SNPs are required for genotyping a population of European or East Asian versus African ancestry. With the supposition that causal variants are shared among different populations, transethnic mapping by taking advantage of differences in LD and allele frequency might be expected to facilitate identification of causal variants. Here, the significantly shorter haplotype blocks in black populations could help to both fine-map association signals identified in European populations and identify novel ethnic-specific signals.³³

Most of the disease-related GWAS loci discovered in Europeans have been replicated in populations of East Asian ancestry. A strong and significant correlation of ORs of specific SNPs for CAD and 27 other diseases across European and East Asian samples has recently been demonstrated, indicating that causal variants, in general, are shared between these 2 populations.³⁴ As would be expected, the SNPs that failed replication in East Asian populations mapped to genomic regions with different patterns of LD. The first published GWAS for CAD, including a substantial number of South Asians, was by the C4D Consortium. Their discovery data set included ≈15 000 CAD cases; half of whom were of South Asian descent.²² Of the 11 previously reported common variants for CAD confirmed in this study, the effect was directionally consistent in European and South Asian populations. However, the OR for several of these, including 9p21.3 and SORT1 (encoding sortilin-1 and related to lipid metabolism), was somewhat lower in the South Asian studies. The recent 1000 Genomes–based meta-analysis included a small number of South Asian (13%) and East Asian (7%) subjects but did not report on ethnic-specific effects.²⁵ In summary, despite smaller data sets, the directional effects of CAD risk alleles identified in European populations on CAD risk in South Asians are generally concordant. In contrast to a similar effect size of CAD risk alleles in East Asian and European populations, for many, the effect size is apparently attenuated in South Asians, possibly not only because of the interaction with unknown genetic or environmental factors but also because the causal variant(s) may be inadequately tagged by markers present on available genotyping arrays, resulting in blunted genetic effects.

In contrast to the findings in East and South Asians, the majority of CAD-associated loci (an exception being rs599839 at the SORT1 locus)³⁵ identified in European populations have failed to replicate or shown considerably reduced effect size in black populations.³⁴ This is despite the high prevalence of CAD among blacks and the potential advantage of interrogating their shorter LD blocks for fine mapping of previously identified CAD loci. Two signals previously reported for CAD at the 9p21.3 locus^36,37 did not achieve replication in independent black data sets, and other signals for CAD have not been identified. In the Population Architecture Using Genomics and Epidemiology (PAGE) study, a consortium of multiancestry, population-based studies, 25% of tag SNPs identified in European GWAS had significantly different effect sizes in black cohorts.³⁸ This might imply limited sharing of causal variants between Europeans and Africans when compared with other ethnicities. Given the lower level of LD in African populations, the index SNP identified in European studies may simply fail to tag potentially shared casual variants.³⁴ However, replication has been achieved for signals associated with discrete CAD risk factors.³⁷ It is possible that the genetic risk of CAD in blacks is related more strongly to the genetic contribution to prevalent risk factors, including hypertension and obesity.³⁹ Given that multiple genetic variants of small effect size are believed to account for much of the heritability of CAD in Europeans, these are likely to be greater in number and more difficult to detect in populations of African descent. Here, it should be noted that the sample sizes of these studies in blacks are of smaller magnitudes than the sample sizes of the studies involving individuals of other ancestries, thus limiting statistical power.

Epidemiological studies have identified a large number of biomarkers that are associated with CAD. However, 2 of the most important problems of observational epidemiology are confounding and reverse causation. These factors limit the ability of these types of studies to identify causality of a biomarker or a risk factor. To circumvent confounding and reverse causation, a growing number of epidemiologists are turning to the Mendelian randomization (MR) method, a method that can be likened to the randomized controlled trial (RCT). This method that combines genetic and observational data takes advantage of the random assortment of genetic variants (=alleles) at conception; therefore, by studying common genetic variants that associate with the exposure variable of interest (eg, low-density lipoprotein [LDL] cholesterol [LDL-C]), reverse causation and most confounding can be avoided. The methodology is based on a simple tenet—if a biomarker has a causal association with a disease, the genetic determinants of the biomarker will also associate with disease risk.⁴⁰ MR has limitations that are shared with RCTs, and one of the most important of these is the lack of pleiotropy, that is, the genetic variant(s) in question must influence only the biomarker of interest. This is not always the case for plasma lipid traits, where pleiotropic effects are evident for several genes, including CETP, LPL, and APOA5. Using MR, major advances have been made in determining the causal associations between plasma levels of lipoprotein(a), LDL-C, and triglycerides (as a marker of remnant cholesterol) and risk of CAD, whereas C-reactive protein (CRP) and high-density lipoprotein (HDL) cholesterol, despite both being robust risk markers, have not been shown to be causal (Figure 2).

Elevated levels of lipoprotein(a) are associated with increased risk of CAD, and common copy number variants and SNPs in the LPA gene, which associate with increased lipoprotein(a) levels, cause an increased risk of CAD^41–43 and of aortic valve stenosis.^44,45 Although the exact mechanisms are not clear,⁴⁶ these data suggest that apolipoprotein(a) is directly causal and therefore a potential drug target.

That elevated levels of LDL-C are causally associated with an increased risk of CAD is evident from familial hypercholesterolemia because of mutations in the LDLR. Genetic variants at the PCSK9, NPC1L1, and HMGCR loci uniquely associate with plasma concentrations of LDL-C and are therefore also predictive of CAD risk. As originally reported by Abifadel et al,⁴⁷ gain-of-function mutations in PCSK9 cause autosomal dominant familial hypercholesterolemia. Cohen et al⁴⁸ later reported that loss-of-function mutations in PCSK9 were associated with reduced plasma concentrations of LDL-C and lower risk of CAD, establishing PCSK9 as a relevant new drug target for lowering LDL-C levels. Ezetimibe reduces LDL-C levels by inhibiting Niemann–Pick C1-like protein1 (NPC1L1), a transporter that in humans is responsible for cholesterol uptake from the intestine into enterocytes and from bile into hepatocytes. Despite this, ezetimibe has only recently been shown to reduce cardiovascular risk.⁴⁹ The Myocardial Infarction Genetics Consortium investigators sequenced NPC1L1 in 7364 CAD cases and 14 728 controls and identified 34 loss-of-function mutations. One of these, p.Arg406X, in a much larger replication study was found to be associated with a 10% lower LDL-C and a 50% decrease in CAD risk.⁵⁰ Similarly, Lauridsen et al⁵¹, in a single-center study of the general population, included 67 385 individuals; of them, 5255 and 3886 developed incident ischemic vascular disease and symptomatic gallstone disease, respectively. Using a genetic score of common variants in NPC1L1, mimicking the effect of ezetimibe, they showed that these variants were associated with lifelong, stepwise reductions in LDL-C levels of ≤3.5%, with a corresponding 18% reduction in risk of CVD and a 22% increase in risk of symptomatic gallstone disease. The study, therefore, suggested that a biologically plausible long-term side effect of ezetimibe treatment might be an increased risk of symptomatic gallstone disease. Finally, Ference et al⁵² in a study of 108 376 subjects in 14 clinical trials reported that common genetic variants associated with each of NPC1L1 and HMGCR conferred a reduction in LDL-C of 2.4 and 2.9 mg/dL and a respective 4.8% and 5.3% lower CAD risk. Because PCSK9, NPC1L1, and HMGCR regulate LDL but do not exhibit effects on other lipid and nonlipid parameters, such studies strongly support a causative relationship between plasma concentrations of LDL-C and CAD. Importantly, multiple variants associated with LDL-C concentrations, that is, in PCSK9^48,53 and NPC1L1^50,51 genes, are more strongly associated with CAD than are associated differences in measured LDL-C or the effect of LDL-C reduction in the statin (RCTs). This is presumably because genetic effects encompass the lifetime exposure to LDL-C (Table 2).^52,54

The role of plasma triglycerides as an independent risk factor of CAD was debated for many years. However, genetic studies have lent credence to the hypothesis that triglyceride-rich lipoproteins or their remnants have a causative role in atherosclerosis.⁵⁵ The major determinant of plasma triglyceride metabolism is lipoprotein lipase, and genetic variants that decrease LPL function confer increased cardiovascular risk^54,56–58 and all-cause mortality.^54,59 LPL activity is increased by ApoA5 and inhibited by ApoC3 and the angiopoietin-like proteins 3 and 4. Loss-of-function mutations in APOC3^60,61 are associated with lower plasma triglycerides and CAD risk, whereas loss-of-function mutations in APOA5 have opposite effects.⁶² In a collaborative analysis, including 20 842 individuals with CAD and 35 206 controls, Sarwar et al⁶³ reported that a common promoter polymorphism (–1131T>C; rs662799) in APOA5 was associated with CAD with an allele-specific OR of 1.18 directionally similar but quantitatively greater than the effects on plasma triglycerides. Similarly, Jørgensen et al⁶⁴ found that genotype combinations of 3 common variants in APOA5 associated with increases in nonfasting triglycerides and calculated remnant cholesterol of 1.10 and 0.40 mmol/L, respectively and with corresponding ORs for myocardial infarction of 1.87 (1.25–2.81).

Essential for successful MR studies is the selection of genetic variants without pleiotropic effects. The major difficulty in studying raised concentrations of triglycerides or remnant cholesterol is the inverse association with HDL cholesterol concentrations.⁵⁵ However, a MR study with genetic variants in several candidate genes that affect the concentrations of remnant cholesterol or HDL cholesterol or both showed that an increase of 1 mmol/L in remnant cholesterol was associated with a 2.8× increased risk of ischemic heart disease that was not attributable to lower HDL cholesterol concentrations.⁶⁵ A recent GWAS also supports that variants associated with high concentrations of triglycerides are causally associated with CAD, after adjustment for HDL cholesterol.⁶⁶ These 2 studies^62,65,66 also showed that genetically low HDL cholesterol was unrelated to cardiovascular risk, in agreement with several previous studies essentially using an MR approach to test causality for HDL cholesterol.^60–65 Because triglycerides can be degraded by most cells but cholesterol cannot, the cholesterol content of triglyceride-rich remnant lipoproteins (remnant cholesterol) is more likely to be the cause of atherosclerosis and cardiovascular disease rather than raised triglycerides. Remnant lipoproteins can, like LDL, enter and be trapped in the arterial intima simply because of their size and possibly via attachment to extracellular proteoglycans.⁵⁵ Taken together, genetic studies strongly support that triglyceride-rich lipoproteins and remnant cholesterol are causal risk factors of CAD and all-cause mortality.

As noted above, MR studies have not supported a direct protective function of HDL cholesterol levels against CAD. Apolipoprotein A-I is the major protein on HDL, but common variants in APOA1 associated with low HDL cholesterol are not associated with an increased risk of CAD.⁶⁷ For the effect of rare variants on CAD risk, the picture is complicated by the fact that some rare variants in APOA1 that associate with low HDL cholesterol, cause amyloidosis, and this, in addition to other symptoms, may manifest as CAD due to small vessel disease, increased intima media thickness, or in later stages as cardiomyopathy. Haase et al^68,69 resequenced the APOA1 gene in >10 000 individuals, genotyped the identified variants in >55 000 individuals from the general population, and reported that the apparent increased risk associated with nonsynonymous mutations in APOA1 became insignificant when variants previously associated with amyloidosis were removed from the analysis. In support, mutations in APOA1 reported to associate with endothelial dysfunction, increased arterial wall thickness, and premature coronary heart disease⁷⁰ have more recently been shown to cause systemic amyloidosis.⁷¹

For other biomarkers, MR studies have provided important insights relevant to drug targets, drug discovery, and potential long-term side effects of drugs. Although high-sensitivity CRP is a robust biomarker for CAD risk and response to stain therapy, neither common genetic variants in the CRP gene associated with substantial increases in CRP levels in the general population⁷² nor rare exonic variants associated with CRP levels⁷³ associate with CAD risk, demonstrating that CRP is not a causal risk factor and hence not a relevant drug target. Similarly, although data from epidemiological and clinical studies supported a potentially important role of lipoprotein-associated phospholipase A2 in atherosclerosis and its sequelae,^74,75 the genetic variants near PLA2G7 and CETP associated with LP-PLA2 levels and activity did not confer CAD risk.^76,77 In accord with these findings, the results of phase III randomized controlled trials of inhibitors of these enzymes (varespladib⁷⁸ and darapladib⁷⁹) were also negative. A closer inspection of MR data before drug design and initiation of clinical trials could, thus, help to mitigate the high costs of drug development. Conversely, 3 of 4 CETP inhibitors either have failed because of adverse events (torcetrapib) or have been stopped because of futility (dalcetrapib and evacetrapib). Nevertheless, common genetic variants in CETP associated with reduced mass and activity are associated with increased HDL cholesterol, smaller decreases in LDL-C and triglycerides and, in the Copenhagen Heart Study, corresponding stepwise reductions in risk of all ischemic end points and all-cause mortality⁸⁰ consistent with other reports.^81–83 However, the effects of small molecule inhibition of CETP are not analogous to the small changes in CETP concentrations because of common genetic variants. Statin therapy lowers CETP by ≈20%,⁸⁴ and as reported by Kuivenhoven et al,⁸⁵ individuals with a common CETP polymorphism associated with higher CETP levels showed reduced progression of atherosclerosis when treated with pravastatin, whereas those with genetically low CETP concentrations did not, suggesting that there may be an optimal window of CETP activity.⁸⁶

Although the effect of the identified susceptibility variants of CAD is individually small, their effects are independent and additive. These can be incorporated into a genetic risk score (GRS) consisting of the number of risk alleles adjusted for their individual effect size. In an analysis performed in 2 prospective cohorts, Whitehall II (WHII; n=5059) and the British Women’s Heart and Health Study (BWHHS; n=3414), individuals in the top versus bottom quintile of an LDL-C GRS consisting of 23 SNPs had a higher risk of CAD (WHII: OR=1.43; BWHHS: OR=1.31).⁸⁷ After adjusting for LDL-C levels, this association was completely attenuated in WHII but not in BWHHS. The value of a GRS for CAD risk prediction beyond conventional risk factors has improved beyond previous studies considering a small number of risk variants^88,89 beyond those related to plasma lipid traits to more recent analyses incorporating a more comprehensive list of recently identified CAD loci.^90–92

Tikkanen et al⁹² constructed a GRS consisting of 28 CAD genetic variants and in 24 124 participants, in 4 population-based, prospective cohorts, determined its association with incident cardiovascular disease events for a mean 12-year follow-up period. Compared with conventional risk factors and family history alone, the GRS improved CAD risk discrimination (C index, 0.856 versus 0.851; P=0.0002). More recently, Mega et al⁹³ evaluated the association of a GRS based on 27 genetic variants with incident or recurrent coronary heart disease, adjusting for traditional clinical risk factors in a community-based cohort study (48 421 individuals in the Malmo Diet and Cancer Study) and 4 RCTs of statin therapy, in primary (Justification for the Use of Statins in Prevention: an Intervention Trial Evaluating Rosuvastatin [JUPITER] and Anglo-Scandinavian Cardiac Outcomes Trial [ASCOT]) and secondary prevention (Cholesterol and Recurrent Events [CARE] and Pravastatin or Atorvastatin Evaluation and Infection Therapy-Thrombolysis in Myocardial Infarction 22 Trials [PROVE IT-TIMI 22]) populations. Their major findings were that the multivariable adjusted hazard ratio for CAD was directly related to the GRS (hazard ratio, 1.72 for individuals in the highest versus the lowest GRS quintile) and that in the statin RCTs, both relative and absolute risk reductions were 3-fold greater in individuals in the highest genetic risk category, with highly significant effects on the number needed to treat. Thus, a GRS can provide important information on future CAD risk beyond traditional risk factors and may have both clinical and economic values in identifying those individuals who will benefit most from the early application of various disease-deferring therapies, including statin treatment.

Overall, the most recent GWAS analysis strongly supports the conclusion that genetic susceptibility to CAD is largely derived from the effect of multiple common SNPs of small effect size.²⁵ However, this analysis was not able to interrogate rare variants present in >1 in 200 individuals, including nonsynonymous mutations in genes of relevance to CAD. Earlier studies that entailed only limited resequencing of candidate regions of the genome in CAD kindreds reported various rare mutations or copy number variants associated with CAD but for most replication has not been achieved.⁹⁴ A 21-bp (7 amino acids) deletion in MEF2A was associated with apparent autosomal dominant CAD in 1 large kindred,⁹⁵ but this was not confirmed in other studies, including the one involving 2 extended families bearing the same copy number variant.⁹⁶

A comprehensive approach to identify and classify coding variants linked to human disease is whole-exome sequencing. This has yielded important findings for Mendelian disease particularly when family members were also available.⁹⁷ It has also shown utility for the genetic diagnosis of various cardiomyopathies and cardiac conduction disorders.⁹⁸ With reference to CAD, the Myocardial Infarction Genetics Consortium Exome Sequencing Project reported on whole-exome sequencing efforts in ≈11 000 individuals. Of note, the identified coding variants were limited to genes relevant to lipoprotein metabolism, including LDLR, APOA5, APOC3, and NPC1L1^50,60,62 and for which there was a large body of pre-existing information. This finding was somewhat disappointing but did provide further credence to the relevance of triglyceride metabolism to atherosclerosis^{55,61,63–65} and the validity of NPC1L1 as a drug target.^49,51,99 Given the perceived need for large sample sizes and subsequent replication, the sample included CAD cases with limited phenotypes, such as acute myocardial infarction without angiographic documentation of disease burden, and controls who may have harbored a burden of nonobstructive atherosclerosis. The inability to follow up on rare variants of high effect size in family members with and without CAD may also have limited the ability to detect novel mutations linked to atherosclerosis. Importantly, studies to date have been significantly underpowered. Recent stimulation analyses indicate that tens of thousands of individuals will be required to confidently detect or exclude rare variant signals at complex disease loci, such as CAD.¹⁰⁰

Despite the recent success of GWASs, close to 80% of the estimated heritability of CAD remains unknown. Some of this missing heritability may be because of gene environment interaction¹⁰¹ where the effect of a given genetic variant is only manifest in the presence of a modifier, such as obesity or cigarette smoking. For example, the effect of common risk alleles for elevated plasma triglycerides was markedly attenuated in lean versus obese individuals.¹⁰² Epistasis is a statistical interaction between ≥2 genetic loci where the effects are nonadditive and has been hypothesized to play an important role in the genetic determination of complex diseases, such as CAD. Epistasis on a genome-wide level has proven challenging because of the requirement for individual genotype data, computational power required for a large number of pair-wise or high-order tests, and the need to correct for multiple testing. However, filtering to consider only those regions with potential biological interaction reduces the necessity for high computational power and the burden of multiple testing correction. We recently reported that SMAD3 is a necessary factor for transforming growth factor-β–mediated stimulation of mRNA and protein expression of type IV collagen genes in human vascular smooth muscle cells and demonstrated a statistical interaction between the COL4A1/COL4A2 locus and the SMAD3 locus by performing pair-wise tests between SNPs at the 2 loci.¹⁰³

It is also likely that many more common variants are linked to CAD but have not achieved genome-wide significance in GWAS because of small effect size or low allele frequency and insufficient sample size. However, on the basis of the premise that clinically informative polymorphisms related to complex disease occur in systems of closely interacting genes,¹⁰⁴ even weakly associated variants may provide important information on the biological basis of disease when such variants cluster within a common functional module or a pathway.¹⁰⁵ Thus, novel insights into the genetic architecture of CAD can be obtained by interpreting genetic findings in the context of biological processes and functional interactions among genes.

A modified version of gene-set enrichment analysis, developed for the interrogation of gene expression data,¹⁰⁶ was designed by Wang et al¹⁰⁷ for the analysis of genome-wide SNP associations, and other gene-set enrichment analysis methods have been more recently applied to GWAS data.^108–112 These analytic algorithms seek to identify a sets of genes whose variants collectively demonstrate strong association with a trait of interest even if the component SNPs individually exhibit relatively modest or nonsignificant association.¹⁰⁵ To identify novel associations between established biological mechanisms and CAD, we recently performed a 2-stage pathway–based gene-set enrichment analysis of 16 GWAS data sets for CAD that included >25 000 subjects with CAD and >66 000 controls¹⁰⁵ using the i-GSEA4GWAS tool¹¹² and the Reactome pathway database.¹¹³ A total of 32 of 639 Reactome pathways tested demonstrated convincing association with CAD. These resided in core biological processes and included pathways relevant to extracellular matrix integrity, innate immunity, axon guidance, and signaling by platelet-derived growth factor, NOTCH, and the transforming growth factor-β–SMAD receptor complex. Importantly, many of these pathways were shown to have strengths of association comparable with those observed in lipid transport pathways.¹⁰⁵

A further extension of the pathway approach is the interrogation of molecular networks that underlie the complex architecture of complex disease.¹¹⁴ A molecular network is based on interactions among diverse molecules, including genes, proteins, and metabolites. These can include interactions between proteins, effects on gene regulation, coexpression, and various functional interactions. In general, genes associated with the same or similar disorders tend to occupy similar molecular networks through physical or functional modules.^115,116 Furthermore, this approach indicates that disease-related genes exhibit network connectivity and network centrality properties that are distinct from other genes.¹¹⁶

Thus, interrogation of molecular networks provides additional information on interactions among gene subsets within a given pathway and highlights potentially important interactions between components of different biological pathways. In the study of Ghosh et al,¹⁰⁵ network analysis of unique genes within the replicated pathways not only further confirmed the known processes, such as lipid metabolism, but also revealed many interconnected functional and topologically interacting modules representing novel associations, for example, the semaphorin-regulated axonal guidance pathway. The axon guidance pathways modulate diverse biological phenomena, including cellular adhesion, migration, proliferation, differentiation, survival, and synaptic plasticity, through the participation of highly conserved families of guidance molecules, including netrins, slits, semaphorins, and ephrins, and their cognate receptors¹¹⁷ and members of these pathways have been highlighted in many recent reports related to macrophage migration,¹¹⁸ expression of inflammatory markers and M2 signals in atherosclerotic plaque,¹¹⁹ and chemokine-directed migration of human monocytes.^120,121 Network centrality analysis further identified genes (eg, NCAM1, FYN, and FURIN) likely to play critical roles in the maintenance and functioning of several of the replicated pathways.

Despite the success of recent GWASs, there has been limited progress in understanding the function of the multiple risk loci identified. Notably, the majority of signals for CAD identified by the GWAS approach are in noncoding regions of the genome. This is not surprising, given that 99% of the genome is nonprotein coding. Nonetheless, ≈10% of this region is under purifying selection,¹²² implying important functional effects. These regions include noncoding RNAs that may influence gene transcription through multiple mechanisms,¹²³ active promoter and enhancers, regions affecting histone acetylation and deacetylation and chromatin remodeling,¹²⁴ susceptibility to DNA methylation, and micro-RNAs and micro-RNA binding sites.

The GWAS-identified polymorphisms are themselves rarely causal but rather in LD with a neighboring or even distal causal polymorphism. Most of the CAD-associated loci identified by GWASs span several kilobases, and some of these may encompass multiple causative variants in ≥1 genes. For common variants, fine-mapping approaches can be limited by large blocks of LD. As an example, the 9p21.3 risk locus encompasses multiple SNPs in tight LD. Dense resequencing failed to identify less common genetic variants with a larger effect size than the original GWAS SNPs.¹²⁵ As reported by the 1000 Genomes Consortium, several CAD loci contain >1 independent signal.²⁵ For example, 7 independent false discovery rate (q<0.05) SNPs were identified at the COL4A1/COL4A2 locus and 5 in the CDKN2BAS (9p21.3) region with 1 uncommon SNP, rs7855162 (MAF, 0.03), exhibiting a higher effect size than the index SNP.²⁵

It is often not fully appreciated that although the association is at the level of genomic DNA, the relevance may be restricted to a particular tissue or organ. An expression quantitative trait locus (eQTL) effect, that is, an effect on mRNA expression on a nearby (cis-eQTL) or distant (trans-eQTL) gene, provides strong evidence for a functional effect of a given SNP. However, recent studies have shown that some eQTLs are shared among tissues, whereas others are tissue specific. Hence, interrogating eQTL databases, for example,¹²⁶ Genotype Tissue expression and ¹²⁷ SNP and CNV Annotation database, requires careful consideration of the cell type or tissue of interest.¹²⁸ Finally, it must be acknowledged that the direct contribution of any given locus could be temporally restricted, for instance, early transient activation followed by sustained epigenetic effects.

The majority of common CAD-associated variants exhibit small individual effects on disease risk, and many are likely to exert their effects by altering promoter or enhancer activity. As a complementary approach to eQTL analysis, public databases, generated by the Encyclopedia of DNA Elements¹²⁹ and Roadmap Epigenomics projects,¹³⁰ genome-wide chromatin profiles of histone modifications, data on transcription factor–binding sites, and chromatin immunoprecipitation sequencing data¹³¹ can be used to identify specific functional regulatory elements.¹³² H3K4me3 denotes active promoters, H3K4me1 denotes enhancers, and H3K4me2, as well as most histone deacetylations, denotes both promoter and enhancer regions.¹³³ DNase I hypersensitive sites are also a mark of open chromatin regions containing promoters and enhancers.¹³⁴ Assay for transposase-accessible chromatin with high-throughput sequencing is a useful method for measuring chromatin accessibility genome wide.¹³⁵ An important caveat is that these analyses require genome-wide chromatin data from a relevant cell type.¹³⁶ These approaches to the identification of regulatory sequences can be followed up by more specific methodologies appropriate to the characteristics of the risk locus in the cell or tissue of interest. Thus, it can first be determined whether a risk locus is likely to harbor functional cis-acting regulatory modules whose activity is altered by an SNP in close LD with a particular risk variant before standard molecular biology approaches in the laboratory (Figure 3).

The poster child for this approach was the identification of a previously unknown role for sortilin-1 in lipoprotein metabolism. Four common GWAS-identified SNPs associated with LDL-C and CAD lie in a noncoding DNA region on chromosome 1p31 between 2 genes, CELSR2 and PSRC1, and downstream of SORT1. In a series of elegant studies, including primary human hepatocytes, Musunuru et al¹³⁷ demonstrated that these SNPs have the strongest hepatocyte eQTL effect for SORT1 and that one of these, rs12740374, creates a C/EBP (CCAAT/enhancer binding protein) transcription factor–binding site and alters the hepatic expression of SORT1 gene, encoding sortilin-1. In further studies in the mouse, genetically increased hepatic sortilin expression was shown to both reduce hepatic APOB secretion and increase LDL catabolism, providing dual mechanisms for the strong association between increased hepatic sortilin-1 expression and reduced plasma LDL-C levels in humans.¹³⁸

In another recent investigation of GWAS significant loci, Sazonova et al¹³⁹ used a complex series of bioinformatic, molecular, cell biology, and system genetic approaches to show that the transcription factor, TCF21, regulates a transcriptional network linking multiple independent CAD loci. The same group demonstrated that TCF21 regulates the development of the epicardial progenitor cells that give risk to smooth muscle cells that contribute to the fibrous cap.¹⁴⁰ Similar approaches have provided new insight into the roles of PHACTR1¹⁴¹ and ADAMTS7¹⁴² in atherosclerosis. ADAMTS7 plays a role in the regulation of vascular smooth muscle cell migration, and Bauer et al¹⁴³ recently demonstrated that Adamsts7^−/− mice were protected from atherosclerosis on an Ldlr^−/− or Apoe^−/− background and showed reduced neointimal formation after femoral wire injury and that Adamsts7^−/− vascular smooth muscle cells showed reduced migration in the setting of tumor necrosis factor-α stimulation, consistent with a proatherogenic effect of ADAMTS7.

The past decade of research has provided a broader understanding of the genetic architecture of CAD and demonstrates that the genetic basis of CAD largely derives from the cumulative effect of multiple common risk alleles individually of small effect size rather than rare variants with large effects on CAD risk. Although traditional risk factors remain important, application of these data using a systems genetics approach has pointed to substantial roles for genes and pathways relevant to vessel wall biology and immune function. A major priority is to apply high-throughput methodology to understand at the molecular and cellular levels the function of each of the novel loci; the majority of which are in noncoding regions of the genome. Beyond functional insight into disease mechanisms, these data have proven clinical utility for interrogation of biomarker causality and drug discovery, through MR, creation of a GRS with predictive power to better identify those individuals who will benefit most from statin therapy.