Introduction

Plasma lipids and lipoproteins are heritable risk factors for coronary artery disease (CAD),¹ with heritability estimates ranging from 40% to 60% for total cholesterol (TC), low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol cholesterol (HDL-C), and triglycerides, and 30% to 60% for CAD.^2–5 Genome-wide association studies (GWAS) for plasma lipids and CAD have successfully identified >95 gene regions for plasma lipids and 46 for CAD.^1,6 Despite the success of GWAS, single-nucleotide polymorphisms (SNPs) at these loci explain only a modest proportion of the heritability—20% to 25% of the heritability for plasma lipids and <10% of the heritability of CAD.¹ These observations have led to the general question of how to account for the unexplained heritability.

Clinical Perspective on p 587

Typically GWAS uses a single-locus test, in which each variant is tested individually for association with a specific phenotype and the best SNP at each locus is reported. A single SNP may not capture the overall amount of variation at a locus because there may be multiple causal variants at a locus. If individual-level genotype data are available, we can detect additional SNPs using conditional analysis; however, conditional analysis is often infeasible when many studies have contributed results to a large-scale meta-analysis.

Recently, Yang et al⁷ developed a conditional and joint association analysis tool that leverages summary-level statistics and estimated linkage disequilibrium from a reference sample with individual-level genotype data. In the current study, we applied this approach to lipid traits studied by the Global Lipids Genetics Consortium (GLGC) as well as to the dichotomous trait of CAD studied by the Coronary Artery Disease Genome-Wide Replication and Meta-Analysis (CARDIoGRAM) Consortium.^5,6

Methods

GWAS Summary-Level Statistics for Lipids and CAD

We obtained the summary-level statistics (rsID, effect allele, the other allele, frequency of the effect allele, effect size, standard error, P value, and sample size) from the GLGC meta-analysis study for lipids⁶ and from the CARDIoGRAM meta-analysis study for CAD.⁵ These studies included ≤100 184 individuals from 46 studies for lipids, and 22 233 cases and 64 762 controls for CAD, respectively.

Reference Samples for Linkage Disequilibrium

We used the European ancestry individual-level genotype of 9796 individuals and phenotype data of the Atherosclerosis Risk in Communities Study (ARIC) cohort⁸ as a reference sample. ARIC represents a large population-based cohort, and this cohort contributed to the GLGC and CARDIoGRAM meta-analyses. SNP quality control was performed, excluding SNPs with missingness >2%, minor allele frequency <0.01 or Hardy–Weinberg equilibrium P value <1×10⁻⁶. Among a total of 805 437 genotyped SNPs, 617 428 SNPs were retained in the ARIC data. We discarded samples with missingness >3% and 1 of each pair of samples with an estimated genetic relatedness >0.25. A total of 8682 individuals of European ancestry in the ARIC cohort were included for linkage disequilibrium calculation. The SNP data for ARIC were phased by MaCH and imputed into the HapMap Phase 2 CEU panel by minimac, the same panel that was used for the initial GWAS.^9,10 We used the best guess genotypes of the imputed SNPs and excluded imputed SNPs with Hardy–Weinberg equilibrium P value <1×10⁻⁶, imputation quality Rsq <0.3, or minor allele frequency <0.01 and retained 2 490 789 SNPs in the ARIC cohort.

Conditional and Joint GWAS Analysis

We performed a stepwise model selection procedure to select independently associated SNPs using the GCTA tool available online (http://www.complextraitgenomics.com/software/gcta/massoc.html) for each lipid trait and CAD. Briefly, the procedure begins with the most significant SNP with P<5×10⁻⁸ in the single-SNP meta-analysis and tests all the remaining SNPs conditional on the selected SNP(s) in the model. It then selects the SNP with the minimum conditional P value and fits all the selected SNPs in a model, dropping the SNP with the largest P value >5×10⁻⁸. The algorithm iterates until no SNP is added to or removed from the model. The joint effects of all selected SNPs are estimated after the model has been optimized. We define a locus as a chromosomal region at which adjacent pairs of associated SNPs are <1 Mb distant. Details about the conditional and joint analysis are fully described elsewhere.⁷

Estimation of the Variance Explained by the Joint Association

We calculated the variance explained using the following equation where β_M and β_J are effect sizes in standard deviation units obtained from the original meta-analysis and the joint analysis, respectively.¹¹

q² = 2 × β_M × β_J × MAF × (1−MAF) × 100

Variance explained was calculated for (1) the top SNPs from original meta-analysis and (2) the top original SNPs plus the additionally associated SNPs found from the conditional and joint analysis.

Replication Analysis for Lipid Traits

We replicated the variance explained by all jointly associated SNPs detected from the GLGC data using 7312 individuals from the Malmö Diet and Cancer (MDC) cohort as an independent sample in 2 models: method A—a multiple regression of the SNPs selected from the discovery set (GLGC); method B—a replication analysis using the SNPs with their effect sizes estimated from the discovery sample (GLGC) to predict the phenotype in MDC. In method A, 2 predictors were created in the MDC cohort by PLINK¹² for each lipid trait, 1 based on all additionally associated SNPs or its proxy SNPs (r²>0.8) and the other based on the GWAS top SNPs only, and the observed lipid phenotypes were regressed on the predictors. In method B, we created 2 predictors in the MDC cohort, but with SNP effects estimated from the GLGC data set, and regressed the observed lipid phenotypes on the predictors. In both methods, adjusted R² values of MDC were compared with the explained variances for lipid traits in GLGC. Also, we checked the independency of multiple associated SNPs within a locus by comparing β from the model using multiple SNPs and β from the model using each SNP within each locus that has the largest number of multiple associated SNPs for each trait.

Informed Consent and Institutional Review Board Approval

Most of the analyses used summary statistics from prior publications. For genetic association analyses in the MDC cohort using deidentified genotype and phenotype data, each participant had provided written informed consent and approval was given by the institutional review board at Partners Healthcare.

Results

Lipid Phenotypes

Using summary statistics of ≈2.5 million SNPs from the GLGC meta-analysis of 100 184 individuals for 4 lipid fractions along with SNP linkage disequilibrium estimated in 8682 unrelated European Americans selected from the ARIC cohort study (see Methods), we identified 62, 61, 68, and 41 jointly associated SNPs for each lipid trait (HDL-C, LDL-C, TC, and triglycerides) with P<5×10⁻⁸ (Tables I–IV in the Data Supplement), respectively. When compared with previous results conducted by conventional conditional analysis in the original GLGC study, we could detect more associated SNPs with each trait (11 versus 29 for HDL-C, 12 versus 26 for LDL-C, 12 versus 24 for TC, and 9 versus 13 for triglycerides; Table V in the Data Supplement).

For the loci where the increasing alleles of at least 2 SNPs were negatively correlated, some associated variants were undetected in the original GWAS. For example, rs180349 and rs3741298 at the APOA1-C3-A4-A5 locus on chromosome 11 did not exhibit a significant association with HDL-C in single-SNP meta-analyses (P value from the single-SNP meta-analysis [P_M]=8.67×10⁻³ and 4.12×10⁻⁴, respectively), but both SNPs reached genome-wide significance when fitted jointly (Table I in the Data Supplement). In addition, the significance and effect size of the leading SNP at the locus also increased (P_M=2.94×10⁻⁴² to P_J=1.84×10⁻⁷³ for rs964184; Table I in the Data Supplement). There were 12 of 33 HDL-C, 10 of 35 LDL-C, 13 of 44 TC, and 8 of 28 triglyceride loci harboring >2 associated SNPs, with the maximum number of 9, 9, 6, and 7 SNPs at a locus, respectively. The lead SNPs (33 for HDL-C, 35 for LDL-C, 44 for TC, and 28 for triglycerides) explained 3.6%, 5.0%, 5.5%, and 5.7% of phenotypic variance, respectively. These values were almost doubled (7.6% for HDL-C, 8.8% for LDL-C, 8.8% for TC, and 8.5% for triglycerides) when all jointly associated SNPs (62 for HDL-C, 61 for LDL-C, 68 for TC, and 41 for triglycerides) were taken into account.

Coronary Artery Disease

Using summary statistics of ≈2.5 million SNPs from the CARDIoGRAM meta-analysis of 22 233 cases and 64 762 controls for CAD along with the same reference SNP data from the ARIC cohort described above, we identified 18 jointly associated SNPs for CAD with P<5×10⁻⁸. Of these SNPs associated with CAD, 3 of 15 loci represent multiple associated SNPs within a single locus (Figure 1; Table VI in the Data Supplement).

• Download figure
• Download PowerPoint

We found 2 association signals for CAD at the 9p21 CDKN2A and CDKN2B locus. We found a significant joint association with CAD in the LDLR region, where multiple common variants for LDL-C and rare mutations in familial hypercholesterolemia have been previously reported.¹³ Two SNPs, rs8099996 and rs1122608, which are 11 024 bp apart, were retained in the stepwise model selection as jointly associated SNPs with P_J<3.5×10⁻¹¹ (Table VI in the Data Supplement). The secondary SNP (rs8099996; P_M=0.67) was masked by the primary SNP (rs1122608; P_M=9.7×10⁻¹⁰) in single-SNP analyses, but it appeared significant (rs8099996; P=5.0×10⁻¹³) in conditional analysis on the primary SNP (Figure 2). This region was also significant for LDL-C and TC (Figure 1C). The APOA5-A4-C3-A1 gene cluster locus was significant for 4 lipid traits as well as CAD (Figure 1A and 1B). When they were fitted jointly, their effects, as well as statistical significance, were substantially increased compared with those in single-SNP analyses. The 15 leading SNPs explained 9.6% of phenotypic variance. The 3 additional SNPs detected by the joint analysis accounted for 1.8% of the variance explained.

• Download figure
• Download PowerPoint

For a set of 184 SNPs in Tables I to IV in the Data Supplement, we evaluated the associations with CAD in CARDIoGRAM meta-analysis. Of these 184 SNPs, 38 SNPs were nominally associated (P_M<0.05) and 20 SNPs showed a significant association after the Bonferroni correction (P_M<2.2×10⁻⁴; Tables I–IV in the Data Supplement).

Replication of the Lipid Results in an Independent Sample

We validated the direction of effect in each SNP between GLGC results and MDC results for all the jointly associated variants (Figure I in the Data Supplement). The multiple regression analysis showed that the prediction R² values of top primary SNPs were 4.6%, 5.8%, 6.7%, and 5.2%, consistent with the estimate of 3.6%, 5.0%, 5.5%, and 5.7% of variance explained in the discovery sample (GLGC), for HDL-C, LDL-C, TC, and triglycerides, respectively (Figure II in the Data Supplement; method A). And the R² values of the additionally associated SNPs were 5.2%, 3.1%, 2.7%, and 2.2%, in line with the estimate of 4.0%, 3.8%, 3.3%, and 2.8% of variance explained by these SNPs in the discovery sample (GLGC), respectively (Figure II in the Data Supplement; method A).

In addition, when we used the SNP effects estimated from the GLGC data set (the second method), the R² values of top primary SNPs were 3.4%, 4.6%, 4.6%, and 4.4%, consistent with the estimate in the discovery sample (GLGC), and those of the additionally associated SNPs were 4.1%, 3.1%, 2.6%, and 2.2%, in line with the estimate of those explained by these SNPs in the discovery sample (GLGC), respectively (Figure II in the Data Supplement; method B). Therefore, these replication analyses in an independent sample confirmed that additional associated variants could explain ≈2% to 4% of phenotypic variation for each lipid trait.

Figure III in the Data Supplement shows that β from the multiple SNP model and β from the single-SNP model are consistent for each locus with the largest number of multiple SNPs (A: HDL-C for CETP locus; B: LDL-C for APOE-C1-C2 locus; C: TC for APOE-C1-C2 locus; D: triglycerides for LPL locus), suggesting that the variants from the multiple SNP model are independent.

Discussion

In this study, we detected several loci where multiple independent SNPs were associated with lipid traits, and accounting for these variants nearly doubled the heritability explained by the previous GWAS results (HDL-C, LDL-C, TC, and triglycerides). In addition, the joint associations of lipid traits were validated in an independent sample.

GWAS results have explained only a fraction of the heritability of complex traits. There has been extensive debate regarding this unexplained heritability, with hypotheses ranging from rare variants to epistasis.^14–16 Here, we explored the possibility of multiple independent signals at a given locus as a contributor to the unexplained heritability.

Conditional analysis has been used as a tool to identify secondary association signals at a locus, starting with the top associated SNP, across the whole genome followed by a stepwise procedure of selecting additional SNPs, one by one, according to their conditional P values. However, nearly always, pooled individual level genotype data are unavailable in large-scale meta-analyses. In that sense, the tool we used is useful in terms of saving computational time and cost because it does not require individual genotype data except for the samples used as a linkage disequilibrium reference. As a result, the current study clearly indicates that an increased portion of the missing heritability could be explained by the joint influence of multiple variants within a locus, suggesting the importance of digging into known loci to identify causal variants and understand the genetic architecture of complex diseases.

For plasma lipids, we found that many signals at specific loci can explain a large proportion of the variance. From our estimation, 9 SNPs in the CETP locus could explain as much as 2.6% of variance explained in HDL-C; 9 SNPs in the APOE-C1-C2 locus explained 2.1% of that in LDL-C; 6 SNPs in the APOE-C1-C2 locus explained 1.1% of that in TC; 7 SNPs in the LPL locus explained 1.7% in that of triglycerides; and 2 SNPs in the CDKN2A/CDKN2B locus explained 2.8% of that in CAD trait. This is consistent with a previous report that suggests greater heritability of common variants in known loci.¹⁷

The variance explained by the top SNPs for each lipid trait from the GLGC data set was relatively small compared with the one from the original article.⁶ This could be because of variability in estimates from different studies as well as the different method used in this analysis compared with that of original GWAS, where only 1 study (ie, the Framingham Heart Study) contributed to the estimation of variance explained. Although we showed evidence of multiple associations at several loci using summary statistics of the CARDIoGRAM meta-analysis, recently published data of the CARDIoGRAMplusC4D meta-analysis with 63 746 CAD cases and 130 681 controls based on the Metabochip array might also be useful to find additional associated signals.¹

In summary, we detected several loci where multiple associated SNPs within a single locus were associated with lipid traits or CAD. For lipid traits, these variants nearly doubled the heritability explained.

Footnotes