Genetic Analysis of 103 Candidate Genes for Coronary Artery Disease and Associated Phenotypes in a Founder Population Reveals a New Association between Endothelin-1 and High-Density Lipoprotein Cholesterol

Study Sample Description

Three independent samples were collected for this study. All three samples comprised individuals with proven French Canadian ancestry (all four grandparents originated from the SLSJ area). All individuals with proven or likely mutations in the low-density lipoprotein (LDL) receptor gene (LDLR [MIM 606945]) and/or lipoprotein lipase gene (LPL [MIM 609708]) (i.e., those with type I or type V dyslipidemia) were excluded from analysis. In LDLR, two deletions (5 kb and >15 kb) and the mutations W66G, E207K, C646Y, C152W, R329X, C347R, and Y468X were screened. In LPL, mutations D9N (rs1801177), N291S (rs268), and P207L were screened. Each individual was extensively phenotyped with a clinical history, a pharmacological report, biometric measurements (height, weight, blood pressure, and waist circumference), and biochemical analysis. Measurements were done after a 3-wk medication washout (for hypocholesterolemic and antihypertensive drugs). The presence of CAD was ascertained on the basis of (1) clinical and electrocardiogram criteria or a positive result of an exercise-tolerance test according to the consensus document of the joint European Society of Cardiology and the American College of Cardiology committee4 or (2) evidence of coronary stenosis of at least 50% in two or more main coronary arteries from coronary angiography for the investigation of ischemic heart disease.

Biochemical Measurements

Blood samples were obtained after a 12-h overnight fast. Biochemical measurements were performed using standard methods on a CX7 Beckman automated analyzer. Apolipoprotein B (apoB) levels were determined using nephelometry. Fasting plasma adiponectin concentrations were measured with a commercial enzyme-linked immunosorbent assay (B-Bridge International).

Gene and SNP Selection

A total of 103 candidate genes were selected on the basis of published evidence of involvement in CAD risk or suspected biochemical contribution to a disease pathway. A strong emphasis was put on lipoprotein metabolism genes (table 1). The gene and SNP panel was specifically designed for the INTERHEART project (see Web Resources) that is currently underway.7 A total of 1,536 SNPs were genotyped for each individual in the family and case-control samples. SNP selection was based on linkage disequilibrium (LD) data from the International HapMap Project,8 previously published functional SNPs, and nonsynonymous coding SNPs. Briefly, SNPs were selected for each gene so as to evaluate all common SNPs (minor-allele frequency [MAF] >0.05) with a r²>0.8 in each of the HapMap population panels (Yoruba of Ibadan, Nigeria; U.S. residents with northern and western European ancestry [CEPH individuals]; and Chinese from Beijing) within 10 kb (both upstream and downstream) of candidate genes. This was done using the LD-select software,9 which operates by selecting a minimal set of markers such that no allele in the data set is correlated at an r²>0.8 with another one. This process was performed separately for each of the three HapMap panels, and a final set of markers was derived from the union of all three population-specific sets. HapMap data release 16 from National Center for Biotechnology Information build 34 was used. Even though the samples used in this study are from a European founder population, tagging SNPs from diverse HapMap populations were genotyped because the INTERHEART project, for which this panel was designed, is a multiethnic study.

In addition, all known common (MAF>0.05 in at least one HapMap population) nonsynonymous coding SNPs were included. Finally, SNPs with substantial evidence of implication in CAD or atherosclerosis on the basis of a review of the current literature were also included. A summary of SNP selection is given in table 2, and a complete list of selected genes and SNPs is available online (see the tab-delimited ASCII file, which can be imported into a spreadsheet, of data set 1 [online only]).

Genotyping

The familial and case-control samples were genotyped using the GoldenGate technology from Illumina. GoldenGate is a proprietary technology based on allele-specific primer extension and highly multiplex PCR with universal primers, as reviewed by Syvanen.10 A fluorescence polarization assay was established to genotype rs5370 (in EDN1) and rs2292318 (in lecithin cholesterol acyl-transferase [LCAT]) in the validation sample (see appendix A for the sequences of primers and probes used).

LD Analysis

LD was analyzed using Haploview.11 Haploview was also used to make graphical representations of LD (i.e., r² values). Haplotype blocks were defined according to the definition given by Gabriel et al.12

Linkage Analysis

Linkage analysis was performed using the MERLIN software package.13 To account for tight LD between SNPs, they were organized in clusters, as described by Abecasis and Wigginton14 (see below for details regarding clusters). Linkage analysis was then performed using these clusters as unlinked markers, thus avoiding inaccuracies in results caused by violation of the intermarker linkage-equilibrium assumption. This is especially important since this effect is more pronounced when parental genotypes are missing, as was the case for much of our sample.15^–17

SNPs with an MAF >0.05 and within 500 kb of each other were clustered together (under the assumption that SNPs >500 kb apart are not in LD and thus would not violate the intermarker linkage-equilibrium assumption), yielding a total of 99 clusters. However, since cluster size was limited to 20 SNPs (a computational limitation), three clusters had to be redefined. To keep the maximum amount of information, no SNP was discarded; instead, these clusters were broken into smaller parts in such a way as to minimize intercluster LD. To accomplish this, two methods were used. First, clusters were broken according to the haplotype block definition of Gabriel et al.12 and were visualized with Haploview.11 Second, an algorithm was developed to determine the “break point” that minimizes pairwise r² between the two newly formed clusters. Both methods yielded similar results. A total of 105 clusters were included in the final analysis.

Dichotomous traits were analyzed for linkage by use of Kong and Cox linear allele-sharing model LOD scores.18 QTL analysis was done using a variance-components method13 for age- and sex-adjusted traits. To gain improved confidence in type I error estimation and to allow for the fact that genome coverage was quite heterogeneous (resulting in decreased power and thus overconservative correction for multiple-hypothesis testing), a gene-dropping experiment was repeated 1,000 times, and empirical P values were evaluated. LOD scores were considered significant when they exceeded a 5% experimentwise type I error cutoff (here, an experiment is the testing of a single trait at all 105 SNP clusters).

Association Analysis

Association analysis was performed on the family sample and the case-control sample separately. Furthermore, a joint analysis (with both samples combined) was also done. Both additive and dominant genetic models were tested.

The family-based association test (FBAT) was performed using QTDT.19 To do so, the identical-by-descent matrix computed using MERLIN was used to infer major additive effects at each tested locus, as required by QTDT. The model used included environmental, polygenic, and a major genetic additive effects in its variance assessment. Because FBAT tests for transmission of disease alleles rather than total association, this analysis is immune to population stratification. “Total association” was also tested in the familial sample. The model used included environmental and polygenic effects in its variance assessment and thus accounted for the intrafamilial quantitative-trait correlation. To make sure that positive results were not the result of population stratification, the “stratification” option of QTDT was used. As expected from the stringent criteria for subject selection, no evidence of stratification was found, thus justifying the use of the “total association” option for our analysis.

The case-control and validation samples were analyzed using standard logistic regression (for CAD status) and anaylsis of variance (for quantitative traits). Analyses were done using R (The R Project for Statistical Computing Web site), an open-source statistical package.

Joint analysis of the familial and case-control samples was performed with QTDT by use of the “total association” option. As for the family sample, the model for the combined samples included environmental and polygenic effects in its variance assessment and thus accounted for the intrafamilial quantitative-trait correlation.

Because variance-components approaches can be sensitive to deviation from normal distributions and because permutation testing is not practical for joint analysis, traits were log-transformed when necessary (triglycerides, high-density lipoprotein [HDL] cholesterol, and adiponectin concentrations), and outliers were manually removed after inspection of normal quantile-quantile plots (<10 observations per analyzed trait were removed).

Since there is a possibility that ascertainment for CAD status might induce false-positive results, all statistically significant associations were checked for interaction with case-control status. In other words, analysis was done conditional on the CAD status, to make sure that effects were the same in cases and controls. No evidence of interaction was observed, and cases and controls were thus pooled together for quantitative-trait analysis (data not shown).

To minimize the number of independent statistical tests performed, the analysis (including linkage analysis) was restricted to CAD status, waist circumference, HDL cholesterol, LDL cholesterol, triglycerides, apoB, and adiponectin plasma concentrations. No environmental exposure was included in our analysis.

Study Samples

A total of 884 individuals from 142 different pedigrees were included in the family sample (with average sibships of size 5.7 individuals). A summary of their characteristics is given in table 3. Only 47 parents were available for genotyping, and 276 individuals had CAD. As expected, the case samples were mostly male and had lower levels of HDL cholesterol and lower levels of adiponectin than did the control samples. Unexpectedly, case samples had lower arterial pressure compared with that of control samples. This may be explained by a stronger adherence to antihypertensive medication among case patients than among controls. Moreover, case patients are expected to receive more aggressive treatment than are unaffected family members. Finally, changes in lifestyle following a cardiovascular event may also explain this observation.

A total of 558 individuals were included in the case-control sample. As expected, the case samples were mostly male and had lower levels of HDL cholesterol, lower levels of adiponectin, and higher waist circumference than did the control samples (table 3). Unexpectedly, case samples had lower levels of triglyceride compared with levels in controls. This may reflect the fact that control subjects were recruited through a lipid clinic. Alternatively, it may result from changes in lifestyle following a cardiovascular event, as well as an unwillingness to discontinue medication for the purposes of the study.

A total of 806 individuals were included in the validation sample. As expected, the case samples were mostly male and had lower levels of HDL cholesterol, higher levels of triglyceride, lower levels of adiponectin, and higher waist circumference than did control samples. As shown in table 3, case subjects also had, on average, lower blood pressure than that of control subjects.

Genotyping

Of 1,536 SNPs genotyped in the familial and case-control samples, 1,481 passed quality-control requirements (55 SNPs failed). Furthermore, our analyses were restricted to SNPs with an MAF >0.05, for a total of 1,179 SNPs. Because some of these SNPs are in perfect LD in our sample, a further 200 SNPs were excluded from analysis. Thus, 979 SNPs were used for the final analysis. Since the panel used was designed to capture common and meaningful genetic variation in diverse populations derived from Asia, Africa, and Europe, it is not unexpected that several minimally or noninformative SNPs will be observed when this panel is used in a single founder population. Using only the CEPH individuals for tagging, we would have included only 879 of the 979 SNPs to capture all information with r²>0.8. The additional 100 SNPs are informative in the Asian and African populations for tagging of haplotypes but are redundant in Europeans (i.e., the LD between two SNPs is >0.8 in Europeans and <0.8 in Africans and/or Asians). These 100 SNPs are common in Europeans (MAF>5%) but are in LD (0.8<r²<1) with a SNP already included in the genotyping panel. Overall, the call rate for the 979 SNPs used in the analyses was >99%. No significant deviation from Hardy-Weinberg equilibrium was observed.

Linkage

A single SNP cluster was deemed to be suggestive of linkage according to the Lander and Kruglyak definition.20 This result was for an HDL cholesterol QTL at the MTHFR/NPPA/NPPB SNP cluster, which had a LOD score of 2.52 and a corresponding P value of .0003. The HDL cholesterol linkage peak at the MTHFR/NPPA/NPPB locus remained statistically significant after the gene-dropping experiment (P<.05). No other significant peak was observed. Complete linkage results are available online (see the tab-delimited ASCII file, which can be imported into a spreadsheet, of data set 2 [online only]).

Association

Testing all SNPs for association with a trait generates a considerable number of tests. To account for multiple-hypothesis testing, a conservative Bonferroni correction was applied, with P values ⩽5.0×10^-5 (.05/979) deemed to be significant. Use of such a threshold results in a type I error of 5% per trait analyzed. Briefly, a total of 12 significant associations were found using the more powerful joint analysis (summarized in table 4). The SNPs at rs5370 (in EDN1), rs3764261 (in CETP), and rs2292318 (in LCAT) were associated with HDL cholesterol levels. The SNPs at rs619054, rs651821, rs662799, and rs5128 (all part of the APOC3/APOA4/APOA5 cluster) were associated with triglyceride levels. Whereas the SNP at rs7412 (in APOE) was associated with both LDL cholesterol levels and apoB levels, the SNPs at rs405509 and rs429358 (also in the APOE locus) were associated only with apoB levels. Finally, the SNP at rs266729 (in the ADIPOQ locus) was associated with adiponectin concentrations. No SNP passed the Bonferroni cutoff in either the case-control or the family sample without also being significant in the joint analysis.

Because the associations between HDL cholesterol and rs5370 (in EDN1) and rs2292318 (in LCAT) are new, these two SNPs were genotyped in an independent validation sample. Whereas the association with rs2292318 was not statistically significant (one-sided P=.12), the association with rs5370 was replicated with statistical signficance, with a one-sided P equal to .004. Furthermore, this latter association was found to be sex dependent, with a much stronger association in women (two-sided P=1.3×10^-5 for the joint analysis; one-sided P=.007 for the validation sample) than in men (two-sided P=.14 for the joint analysis; one-sided P=.07 for the validation sample). The call rate for rs5370 in the validation sample was 98.8%, and the call rate for rs2292318 was 99.2%. Both SNPs were in Hardy-Weinberg equilibrium. An overview of the association analysis is given in figure 1.