Common genetic variation in multiple metabolic pathways influences susceptibility to low HDL-cholesterol and coronary heart disease.

A low level of HDL-C is the most common plasma lipid abnormality observed in men with established coronary heart disease (CHD). To identify allelic variants associated with susceptibility to low HDL-C and CHD, we examined 60 candidate genes with key roles in HDL metabolism, insulin resistance, and inflammation using samples from the Veterans Affairs HDL Intervention Trial (VA-HIT; cases, n = 699) and the Framingham Offspring Study (FOS; controls, n = 705). VA-HIT was designed to examine the benefits of HDL-raising with gemfibrozil in men with low HDL-C (≤40 mg/dl) and established CHD. After adjustment for multiple testing within each gene, single-nucleotide polymorphisms (SNP) significantly associated with case status were identified in the genes encoding LIPC (rs4775065, P < 0.0001); CETP (rs5882, P = 0.0002); RXRA (rs11185660, P = 0.0021); ABCA1 (rs2249891, P = 0.0126); ABCC6 (rs150468, P = 0.0206; rs212077, P = 0.0443); CUBN (rs7893395, P = 0.0246); APOA2 (rs3813627, P = 0.0324); SELP (rs732314, P = 0.0376); and APOC4 (rs10413089, P = 0.0425). Included among the novel findings of this study are the identification of susceptibility alleles for low HDL-C/CHD risk in the genes encoding CUBN and RXRA, and the observation that genetic variation in SELP may influence CHD risk through its effects on HDL.

Caucasian race (n = 58) (4 missing information). Data used in our statistical analyses were obtained at baseline, before treatment with gemfi brozil. The characteristics of the cases at baseline, as well as basic demographic information, are provided in Table 1 .
The Framingham Offspring Study is a longitudinal study designed to identify risk factors for cardiovascular disease in subjects from Framingham, MA ( 16 ). A total of 866 males from FOS were selected for use as controls in our case-control analysis. Controls were free of CHD, at least 39 years old at exam six (data collected between 1995 and 1998), and self-reported Caucasian. Plasma lipid and demographic information for the controls at exam six are provided in Table 1 .
All participants in both cohorts provided written informed consent. The VA-HIT study was approved by the Human Rights Committee of the Cooperative Studies Program Coordinating Center and by the Institutional Review Board (IRB) of each VA medical center. Study protocols were approved by the IRB for human research at Tufts University, Boston University, and the Broad Institute.

Biochemical analyses
Plasma lipid concentrations were determined from blood samples collected from subjects, after a 12-14 h fast, into tubes containing 0.1% EDTA. Plasma was isolated and frozen for subsequent analysis of plasma lipid, lipoprotein, and apolipoprotein (apo) concentrations. Plasma total cholesterol and triglyceride concentrations were determined using enzymatic assays ( 17 ). Plasma HDL-C concentrations were measured after dextran sulfatemagnesium precipitation of apoB-containing lipoproteins ( 18 ). LDL-C levels were calculated with the equation of Friedewald et al. ( 19 ).

SNP selection and genotyping
Because HDL-C is a complex trait, we considered it important to explore common variation not only in genes whose products directly infl uence HDL metabolism but also in genes whose products may indirectly infl uence HDL concentrations. Thus, candidate genes included those associated with diabetes and/or insulin resistance, as well as those involved in infl ammatory pathways. As illustrated in Fig. 1 , many of the candidate genes have been associated with more than one phenotype (i.e., LIPG with lipoprotein metabolism and infl ammation). For each candidate gene, the entire coding length of the gene (from the fi rst 5 ′ and 3 ′ UTR) plus 10 kb upstream and downstream of the gene were downloaded from HapMap for the CEU sample (Genome Build 36, and Version 126 of dbSNP), allowing us to search for variation in the promoter and other regulatory regions. SNPs were selected using the Tagger algorithm (www.broad.mit.edu/ mpg/tagger) to identify SNPs (tagSNP) within each gene having a pair-wise r 2 > 0.9 ( 20 ). Priority was given to SNPs that had higher design scores for Illumina genotyping. A list of all the candidate genes and number of tagSNPs selected in each gene is presented in supplementary Table I. To address potential population structure, 274 ancestry informative markers (AIM) were selected for genotyping (21)(22)(23). Inclusion criteria for AIMs included 1) not within 60 base pairs (bp) of another AIM SNP; 2) not in a candidate gene; 3) a minor allele frequency > 0.05; 4) an Illumina design score > 0.6; and 5) an appropriate Illumina validation class.
Genomic DNA was extracted from whole blood samples using either QIAamp mini kits (Qiagen) or Generation Capture Column kits (Gentra Systems). DNA isolated from the case samples was of low concentration and was whole-genome-amplifi ed (WGA) to provide suffi cient DNA for genotyping. For consistency, Environmental and metabolic factors that are commonly associated with low HDL-C concentrations include alcohol consumption, dietary saturated fat, decreased exercise, cigarette smoking, obesity, and diabetes (7)(8)(9). Low HDL-C is commonly seen in association with the metabolic syndrome, which is characterized by visceral obesity, elevated glucose and blood pressure, dyslipidemia, and frequently, insulin resistance and a pro-infl ammatory state ( 10 ). In addition to environmental factors, strong evidence also exists for the role of genetics in the determination of HDL-C levels. HDL-C level is a heritable characteristic, with heritability estimates in the range of 40-60% ( 11 ). Most of the variation in HDL-C observed at the population level is of multifactorial origin and is the result of the complex interaction between genetic and environmental factors.
In the present study, we utilized a candidate gene approach to extend our earlier work (12)(13)(14) with cases recruited through the Veterans Affairs HDL Intervention Trial (VA-HIT) and controls selected from the Framingham Offspring Study (FOS). VA-HIT was designed to examine the benefi ts of HDL-raising with gemfi brozil in men with low HDL-C ( р 40 mg/dl) and prevalent CHD ( 15 ). To identify susceptibility loci for the low HDL-C/CHD trait, we examined the allelic variation of 1,114 singlenucleotide polymorphisms (SNP) in 60 candidate genes using cases from VA-HIT before treatment and controls selected from the FOS. We hypothesized that susceptibility to the low HDL-C/CHD trait would be associated with common variants derived from genes involved in HDL metabolism, insulin resistance, and infl ammation. We further compared the cases to controls having an HDL-C of < 40 mg/dl, as well as to controls having an HDL-C of у 40 mg/dl. These comparisons allowed us to evaluate whether the genetic variants were infl uencing CHD through HDL-C levels or via alternative mechanisms.

Study design
The rationale, design, and methods for VA-HIT have been described elsewhere in detail ( 15 ). Briefl y, men were recruited at 20 Veterans Affairs (VA) medical centers throughout the United States. Eligibility for the trial required a documented history of CHD (defi ned as a history of myocardial infarction, angina corroborated by objective evidence of ischemia, coronary revascularization, or angiographic evidence of stenosis greater than 50% of the luminal diameter in one or more major epicardial coronary arteries); age < 74 years; HDL-C level р 40 mg/dl (1.0 mmol/l); LDL-C level р 140 mg/dl (3.6 mmol/l); plasma triglyceride concentration р 300 mg/dl (3.4 mmol/l); and absence of coexisting conditions (cancer, excluding skin; other life-threatening diseases, such as chronic pulmonary or kidney disease; alcohol or substance abuse; cholelithiasis; or marked left ventricular dysfunction). Information on age, alcohol consumption, smoking status, blood pressure, body mass index (BMI), and diabetes was available for all subjects enrolled in VA-HIT. Based on the existence of informed consent for genotyping analysis, 870 subjects from the VA-HIT study were available for the present study as cases. Subjects were self-identifi ed as Caucasian (n = 808) or non-AIMs) met these quality control criteria and were used in our analysis.

Assessment of population stratifi cation
Population structure, or stratifi cation, is the presence of a systematic difference in allele frequencies between subpopulations in a sample, possibly due to differences in ancestry. Population structure can cause bias in the results of an association study and can lead to false positive fi ndings. The smartpca package in the software EIGENSTRAT ( 24,25 ) was used on the AIMs to determine population structure within the sample. Smartpca performs effi cient principal components analysis (PCA) on the genotype matrix and provides principal components (PC) that describe the genetic variability in the sample. A modifi ed PCA ( 26 ), described below, was conducted on the study sample plus the four HapMap samples [Utah residents with Northern and Western European ancestry from the CEPH collection (CEU); Yoruba in control DNA samples underwent WGA as well. Genotyping was performed using Illumina GoldenGate Bead Array platform with a 1536-SNP OPA (GS0010268-OPA). Sentrix Array Matrix (SAM) fi ber-optic bundles manufactured at Illumina using two lots of oligo-beads (beadtype v4 or v5) were used to image the genotypes. The use of different beadtypes is not expected to affect the performance of SNP assays. No duplicated samples were processed on both beadtypes, but allele frequencies were similar between beadtypes. Individuals with a call rate of < 90%, missing phenotype information, or у 50% genotype data inconsistent with previous genotyping were omitted from analysis (supplementary Table II). A total of 1,404 individuals (cases, n = 699; controls, n = 705) met quality control criteria and were included in the fi nal analysis. SNPs were excluded based on 1) call rate of < 0.90; 2) minor allele frequency (MAF) < 0.05; or 3) Hardy Weinberg Equilibrium (HWE) P value in the control sample < 0.001. A total of 1,355 SNPs (1,114 candidate gene SNPs and 241 Values are either mean ± SD or n (%). a A two sample t -test was used to obtain the P value for continuous variables, and a chi-square test was used to obtain signifi cance for categorical variables to compare cases (n = 699) to all controls (n = 705). b Cannot perform test due to missing information. Candidate genes selected for analysis. Because HDL-C is a complex trait, we explored common variations in genes whose products directly infl uence HDL metabolism, as well as in genes whose products may indirectly infl uence HDL concentrations (diabetes, insulin resistance, and infl ammation). via alternative mechanisms. The fi rst subgroup analysis provided a comparison of those with low HDL-C and CHD to those who have neither, whereas the second analysis provided a comparison of those who have CHD versus those without CHD among men with low HDL-C. The intra-cohort comparison within FOS allowed us to identify SNPs associated with HDL-C levels. Three covariate adjustments were made: 1) age and PC1-4 (labeled "MinAdj"); 2) age, PC1-4, body mass index (BMI), systolic blood pressure (SBP), diabetes status, smoking status, and alcohol use ("MultiAdj"); and 3) age, PC1-4, BMI, SBP, diabetes status, smoking status, alcohol use, HDL-C, triglycerides, and total cholesterol ("LipidAdj"). Adjusting for lipid levels allowed us to compare CHD cases to controls, after controlling for the differences in lipid levels. We note that differences in unmeasured exposures are always a concern and recognize this as a limitation of our work. All analyses used logistic regression in PLINK ( 28 ) with a general genetic model. A general model was used to avoid power loss due to mis-specifying the genetic model ( 29 ). We did not account for family structure in the association analyses, because only extended familial relationships (i.e., cousins) existed among some of the control subjects.
Statistical signifi cance was determined using an empirical P value (AdjP) obtained by a permutation strategy ( 30 ) to adjust for testing multiple SNPs within each gene. A null distribution of minimum P values was generated based on permuting the outcome phenotype (keeping the corresponding covariate values) and performing the analysis as described above. We did not permute the genotypes to maintain the linkage disequilibrium within the genes intact. The minimum P value across all the SNPs in the same gene was documented for 10,000 permutations of the phenotype and compared with the observed P value. This approach provides correction for testing multiple SNPs within each gene. To take into account multiple outcomes, we only conducted subgroup analysis for SNPs that were signifi cant in the primary analysis, as an approach to refi ne the association. False discovery rate (FDR) ( 31 ) was calculated in SAS based on the 60 genes (1,114 candidate gene SNPs) analyzed using the observed P values.
SNP interaction with age analyses were conducted with SNPs that were signifi cantly different between cases and controls, after adjustment for multiple testing within each gene ( P Adj < 0.05), and in silico function predictions were obtained ( 32 ).

RESULTS
Baseline characteristics of the cases (VA-HIT) and controls (FOS) are presented in Table 1 . Signifi cant differences in baseline characteristics were observed between the cases (n = 699) and controls (n = 705). Cases tended to be older and have higher BMI; there are a greater percentage of diabetic and hypertensive subjects relative to controls. Consistent with the selection criteria of VA-HIT, cases also had lower plasma concentrations of total, LDL-, and HDL-cholesterol and higher triglyceride levels compared with controls. Cases were also more likely to be on drug therapy (diabetes medication or antihypertensive medication) than their control counterparts. Note that statistical adjustments were made to account for the differences in these measures when testing for associations between each candidate gene and case status. Table 2 shows the SNPs that were significantly different between cases and controls, after adjustment for multiple testing within each gene ( P Adj < 0.05). SNPs signifi cantly associated with case status were located in the Ibadan, Nigeria (West Africa) (YRI); Japanese in Tokyo, Japan (JPT); and Han Chinese in Beijing, China (CHB)] using the selected AIMs. Plotting PC1 versus PC2 shows that some subjects in our sample are near the YRI individuals ( Fig. 2 ). Subjects clustering near the YRI sample (PC1 > 0.03), denoted with a nonCaucasian race, or removed based on outlier detection (see below), were excluded from analysis to obtain a homogeneous sample of Caucasian individuals for analysis (supplementary Table II). The rationale for excluding these subjects is based on the fact that, due to a lack of control subjects having a PC1 of > 0.03, we cannot adequately adjust for the nonCaucasian subjects in the cases. Including multiple population groups in a sample without appropriate adjustment leads to an increase in Type I error ( 27 ); therefore, only Caucasian individuals were included in the analyses.
A modifi ed PCA on the study subjects was performed ( 26 ). A few family members in a sample are known to bias a PCA, and the entire control sample may have extended family members. To take this into account, we performed PCA on the case sample and a subset of unrelated controls based on the full Framingham pedigree fi le, and then inferred the PCs for the remaining individuals based on the SNP weights. Thirty-three subjects were removed during EIGENSTRAT's default outlier removal process, which iteratively performs PCA, removing subjects that are more than six standard deviations from the mean for the top ten PCs over fi ve iterations.

Statistical analyses
The primary outcome was the combined trait of low HDL-C/ CHD. Low HDL-C was defi ned as HDL-C < 40 mg/dl. Subgroup analysis included comparing 1) low HDL-C/CHD and control subjects with HDL-C < 40 mg/dl; 2) low HDL-C/CHD and control subjects with HDL-C у 40 mg/dl; and 3) control subjects with HDL-C < 40 mg/dl and control subjects with HDL-C у 40 mg/dl. These comparisons allowed us to evaluate whether the genetic variants were infl uencing CHD through HDL-C levels or In addition to comparing CHD cases to all control subjects, we conducted three subgroup analyses, in attempt to determine if the SNP associations with low HDL-C/CHD were acting through the HDL pathway or via other mechanisms. We compared 1) cases (n = 699) to controls with a HDL-C у 40 mg/dl (n = 435); 2) cases to controls with a HDL-C < 40 mg/dl (n = 270); and 3) controls with a HDL-C < 40 mg/dl to controls with a HDL-C у 40 mg/dl. Subgroup analysis was only conducted for SNPs that reached the signifi cance threshold in the primary analysis. The results of the subgroup analyses are provided in Tables 3-5 . All SNP associations identifi ed in the primary analysis remained signifi cant when comparing cases to controls having a HDL-C у 40 mg/dl. A 0.10 difference in minor allele frequency between cases and controls was observed for SNP rs4775065 in LIPC , with the minor allele having a detrimental effect. Relative to the major allele homozygous genotype, an increased odds ratio (OR) for low HDL-C/CHD was seen in heterozygotes (OR = 2.36; 95% CI: 1.83-3.05), while an OR of 1.54 (95% CI: 0.82-2.89) was observed in minor allele homozygotes. The association remained signifi cant after adjustment for plasma lipid levels ( P = 1.15E-06). When controls with a HDL у 40 mg/dl were compared with controls with a HDL-C < 40 mg/dl, we did not fi nd rs4775065 genes encoding hepatic lipase ( LIPC ; rs4775065; P < 0.0001); cholesteryl ester transfer protein ( CETP ; rs5882; P = 0.0002); retinoid X receptor ␣ ( RXRA ; rs11185660; P = 0.0021); ATP-binding cassette transporter A1 ( ABCA1 ; rs2249891; P = 0.0126); ATP-binding cassette transporter C6 ( ABCC6 ; rs150468 and rs212077; P = 0.0206 and P = 0.0443); cubilin ( CUBN ; rs7893395; P = 0.0246); apolipoprotein A-II ( APOA2 ; rs3813627; P = 0.0324); P-selectin ( SELP ; rs732314; P = 0.0376); and apoC-IV ( APOC4 ; rs10413089; P = 0.0425). All SNPs in Table 2 were in HWE, and similar results were seen with the minimal adjustment model (results not shown). None of the reported SNPs had a signifi cant interaction with age (all interaction P values > 0.37). A model fi t with the ten reported variants for the primary outcome of low HDL-C/CHD indicated a generalized-R 2 of 0.14, indicating a 14% improvement over a null model for low HDL-C/CHD, with an area under the curve (AUC) of 0.687. Addition of risk factors (age, BMI, SBP, diabetic status, smoking status, alcohol consumption) to the ten reported variants provides a 37% improvement over the null model with an AUC of 0.810, while a model with just risk factors has a 24% improvement over the null model and an AUC of 0.751. Note, in this context a generalized-R 2 should not be confused with the percent of variability explained in ordinary least square regression.  trols with high HDL-C ( P = 0.205). Similarly, patterns in associations for SNPs in CETP (rs5882) and ABCC6 (rs150468 and rs212077) indicate that these SNPs may also infl uence CHD risk through their effects on HDL.
For the SNPs in ABCA1 (rs2249891), CUBN (rs7893395), APOA2 (rs3813627), and APOC4 (rs10413089), the infl uence of HDL-C on CHD status is less clear. The associations observed between case status and ABCA1 and APOC4 indicate a heterozygote genotype advantage, while the association with APOA2 indicates a heterozygote disadvantage. In silico functional predictions ( 32 ) were seen for the SNPs in APOA2 and APOC4 , indicating that these SNPs may regulate transcription by affecting transcription factor binding sites. However, as these fi ndings may be due to gene-gene, allelic, and/or gene-environment interactions, further biological support is required to confi rm these results.
Two SNPs in ABCC6 , rs150468 and rs212077, reached our signifi cance threshold. In the controls, the linkage disequilibrium (LD) measure (D ′ ) between the two SNPs is 0.9 and the r 2 is 0.77, indicating that these two SNPs are capturing the same variant associated with case status. The to be signifi cant ( P = 0.886). However, in comparing cases to controls with low HDL-C, we observed an increase in the odds ratio, indicating that our top SNP in LIPC may not be acting through the HDL pathway to modify CHD risk. Likewise, the RXRA SNP (rs11185660), which is signifi cantly associated with HDL-C level ( P = 0.0079), does not appear to modify CHD risk through the HDL pathway but infl uences both CHD and HDL-C levels.
For rs732314 in SELP , the minor allele also had a detrimental effect. An OR of 1.71 (95% CI: 1.28-2.30) for low HDL-C/CHD was seen for heterozygous carriers of the minor allele versus major allele homozygotes, while an OR of 1.43 (95% CI: 1.01-2.01) was observed in minor allele homozygotes. The effect was enhanced when cases were compared with controls with high HDL-C (OR het = 1.9; OR hom = 1.76) and diminished when cases were compared with controls with low HDL-C (OR het = 1.52; OR hom = 0.99, respectively). These data suggest that rs732314 in SELP may be infl uencing CHD via the HDL pathway. After adjustment for lipid levels, the association between case status and rs732314 was not signifi cant ( P = 0.1695) nor was the association comparing controls with low HDL-C to con- with cases having a minor allele frequency of 0.33 compared with 0.23 for controls. LIPC has consistently been associated with HDL-C in genome-wide studies of lipid traits (34)(35)(36)(37)(38)(39) and has also been implicated as a susceptibility locus for CHD ( 40 ). The lack of a signifi cant association between LIPC variants and HDL-C in the present study may be due to the limited sample sizes in our subgroup analyses, the dichotomizing of HDL-C levels, or the fact that LIPC may be infl uencing CHD risk via a nonHDLrelated mechanism.
We also observed a signifi cant association between case status and rs5882, a nonsynonymous SNP (I405V) in the CETP gene. The minor allele of this variant has been associated with reduced CETP activity, increased lipoprotein particle size, and exceptional longevity in Ashkenazi Jewish probands ( 41 ). Consistent with this fi nding, the minor allele had a protective effect in the present study and appears to be infl uencing CHD through via its effects on HDL. The effect was enhanced when cases were compared with controls with high HDL-C (OR het = 0.62; OR hom = 0.33) and diminished when cases were compared with controls with low HDL-C (OR het = 0.66; OR hom = 0.49).
Included among the novel fi ndings of this study are the identifi cation of susceptibility alleles for low HDL-C/CHD in the genes encoding CUBN (rs7893395) and RXRA (rs11185660), with the minor allele having a protective effect in each case. The cubilin gene encodes a high affi nity receptor for endocytosis of HDL-C and lipid-poor apoA-I ( 42 ), while RXRA has been implicated in altering the risk for Alzheimer's disease through its effects on cholesterol metabolism ( 43 ). In silico functional predictions ( 32 ) were not found with the associated CUBN (rs7893395) or RXRA (rs11185660) SNPs or with SNPs in linkage disequilibrium (>0.7) with the associated SNPs.
Although genetic variation in P-selectin ( SELP ) has previously been associated with atherosclerotic risk in European-American adults from the Coronary Artery Risk Development in Young Adults (CARDIA) Study ( 44 ), we provide new information that SELP may infl uence CHD risk through its effects on HDL-C. When cases with low LD for these SNPs in cases is similar to that seen in controls. A model was fi t jointly with both ABCC6 SNPs, and it was found that neither SNP was signifi cant. We encountered colinearity issues when both SNPs were fi t in the model, with a variance infl ation factor (VIF) of 3.5. Logistic regression values with a VIF greater than 2.5 indicate colinearity. Therefore, a haplotype analysis of the two ABCC6 SNPs (rs150468:rs212077) was performed. We found that the haplotype with alleles C G (frequency of 16%) was signifi cantly different from the haplotype with alleles A C (frequency of 81%) ( P value = 8.12E-05). The ABCC6 haplotype A C had a 56% increase risk of CHD/ low HDL-C compared with haplotype C G.

DISCUSSION
Nearly half of patients with CHD have low HDL-C as their primary lipid abnormality. Strong evidence exists for the role of genetics in the determination of HDL-C levels, with approximately 50% of the variation in HDL-C levels in humans genetically determined ( 11 ). However, HDL-C is a complex genetic trait, which involves the action of multiple genes that interact with each other and with environmental factors. Therefore, we examined 60 candidate genes with roles in HDL metabolism, insulin resistance, and/or infl ammation to identify allelic variants associated with susceptibility to low HDL-C and CHD. Candidate gene studies are cost-effective, with little loss of power for detecting common variants ( 33 ). Moreover, a candidate gene approach has inherent biological plausibility for the associations uncovered.
The majority of susceptibility alleles for the low HDL-C/ CHD trait identifi ed in our study were located in genes with established roles in HDL metabolism. For instance, we identifi ed susceptibility alleles for low HDL-C and/or CHD risk in the genes encoding LIPC (rs4775065), CETP (rs5882), ABCA1 (rs2249891), CUBN (rs7893395), and APOA2 (rs38136627). The strongest association with case status involved a variant in the LIPC gene (rs4775065), portunity to identify genetic variants associated with CHD and low HDL-C. With this cohort, we used a candidate gene approach to identify novel allelic variants in the genes encoding CUBN and RXRA that are associated with susceptibility to the low HDL-C/CHD trait, and we provide new information that genetic variation in SELP may infl uence CHD risk through its effects on HDL-C. We also replicated data from previous studies through the identifi cation of susceptibility alleles for low HDL-C and/or CHD risk in the genes encoding LIPC , CETP , ABCA1 , and ABCC6 . The fi ndings of the present study may help to identify individuals at increased risk for the development of CHD.
HDL-C and CHD were compared with controls with HDL-C < 40 mg/dl, the association between SELP and case status diminished ( P = 0.0403). Specifi cally, the odds ratio for the comparison of minor allele homozygotes to major allele homozygotes was reduced from 1.43 ( Table 2 ) to 0.99 ( Table 4 ) when cases were compared with all controls versus controls with HDL-C < 40 mg/dl, respectively. Conversely, when we compared cases to controls with HDL-C у 40 mg/dl, the association was enhanced (OR hom = 1.9, OR het = 1.76, P = 0.0007). Finally, when we compared controls having a HDL-C < 40 mg/dl to controls having a HDL-C у 40 mg/dl ( Table 5 ), no association between HDL status and the SNP in SELP was observed ( P = 0.2050). Since the effect increased when low HDL-C/CHD cases were compared with healthy controls (HDL у 40 mg/dl), this suggests that the variant in SELP may be associated with CHD via HDL or that having the minor allele for provides a higher risk of having both CHD and low HDL-C. It is also important to note that the SELP SNP associated with increased risk for case status in our study (rs732314) is in strong linkage disequilibrium with rs2235302, a SNP shown to be signifi cantly associated with both soluble P-selectin and carotid intima-media thickness among European-Americans in CARDIA ( 44 ). Finally, in silico analysis demonstrated a functional prediction (transcription factor binding site) for rs732314 in SELP , suggesting that this SNP is functionally relevant.
To assess whether any of the top risk loci identifi ed in the present study may be related to survival, we determined whether the genotype distribution of the reported SNPs varied by decade of age. We found no evidence of a survival bias in our study. Therefore, the SNPs identifi ed in this study do not appear to be related to survival time but, rather, are likely associated with the development of the combined trait of low HDL-C/CHD. Given that the inability to evaluate survival time is a limitation of our study, further research is required to confi rm this fi nding.
There are additional limitations to our study. The reported novel fi ndings are preliminary in nature. To fully confi rm our fi ndings, the results need to be replicated in independent samples and validated through experimentation. Although we adjusted for potential confounding variables in our analyses, we cannot account for potential differences in unmeasured exposures between the cohorts. Furthermore, this study is based on the common variant-common disease hypothesis, which does not allow us to evaluate the contribution of rare alleles to low HDL-C/CHD risk ( 45 ). Finally, while we have replicated the fi ndings of some earlier studies ( LIPC , CETP , ABCA1 , APOA2 , ABCC6 , and APOC4 ), other fi ndings were not replicated in the current study (i.e., APOA5 , LPL , LCAT , LIPG , and PON1 ). APOA5 and LPL have been associated with HDL-C in European and multi-ethnic samples using the Illumina cardiovascular disease (CVD) BeadChip (45)(46)(47). The lack of replication may be due to the relatively modest sample size of the present study or to the distinctiveness of the cohort studied.
The VA-HIT cohort, composed of men with low HDL-C, normal LDL-C, and established CHD, affords a unique op-