CpG-set association assessment of lipid concentration changes and DNA methylation

Lipid levels can be influenced by drug therapy or lifestyle factors such as diet, physical activity, alcohol consumption, and smoking [1]. Lipid levels are also associated with inherited genetic variants (single-nucleotide polymorphisms [SNPs]), as revealed by several genome-wide association studies [2]. However, DNA sequence variation explains only a small proportion of lipid-level variance [2]. Epigenetic modifications (eg, DNA methylation) alter DNA accessibility and hence can be involved in regulating patterns of gene expression. Through regulation of lipid levels, epigenetic mechanisms may contribute to cardiovascular risk profiles [3, 4]. Irvin et al. [3] identified strong association of 4 cytosine-phosphate-guanine (CpG) sites within the CPT1A gene on chromosome 11 with both triglycerides (TGs) and very-low-density lipoprotein C (VLDL-C). Because their analysis examined phenotype associations at each CpG site, a substantial correction for multiple testing was required.

This study’s goals were to investigate region-based associations between DNA methylation sites and lipid-level changes in response to the treatment with fenofibrate in the real data set provided by the GAW20 and to investigate whether improvements in power could be obtained by taking into account correlation between DNA methylation at neighboring CpG sites. We conducted 2 complementary block-based association tests that simultaneously accommodated all CpG sites that fell within a genomic region (block) for the purpose of investigating whether taking into account the correlation between DNA methylation at neighboring CpG sites and reducing the number of tests can improve power.

Turgeon et al. [5] have proposed “principal components of explained variation (PCEV).” The PCEV approach integrates, simultaneously, an optimal data-dimension-reduction technique with testing for association. It provides analytical and empirical p value calculations for testing association between a set of correlated variables (eg, methylation profiles of a genomic region) and 1or more variables of interest (eg, high-density lipoprotein [HDL]/TG).

We contrast PCEV with a variance components (VC) score test method. This is a sequence kernel association test (SKAT)-type association test that decomposes the total variance of a phenotype (eg, HDL/TG) into the variance explained by a block/region-methylation profiles and a residual variance term [6]. Specifically, the model assumes that the phenotypic similarity between subjects is captured by the region–methylation similarity. The VC-score approach significantly reduces the model degrees of freedom compared to standard multivariate regression models.

Suppose we have observed data {Y, x, C} where Y is an n × p matrix of n subjects for which a block of p variables/phenotypes are measured (eg, methylation values at a genomic region/gene with pCpG dinucleotides), x is an n × 1 vector of an observed trait of interest (eg, HDL phenotype) and C is an n × r matrix where the columns are known confounding factors (eg, age, sex).

In this study, we investigated associations between DNA methylation and lipid-level changes in response to fenofibrate treatment in the GAW20 real data set. The smallest p value found by both the VC-score test for TG differences and the single-CpG association test was at gene LGALS9C. However, as a result of multiple-testing power-loss issues, no single-CpG test passed a Bonferroni-corrected threshold. Simulation studies are necessary for more rigorous power assessments comparing region-based and univariate methods. Furthermore, we did not replicate the results of Irvin et al. [3] for baseline lipids, although this is not surprising as we analyzed lipid-level changes. We focused on linear relationships; however, nonlinear association methods may be more favorable/powerful when the primary interest is lipid-level changes.

In all analyses, we adjusted for available/known confounders and for unknown confounders using 4 PCs calculated based on 2000 CpG probes selected randomly from available DNA methylation on all chromosomes. We also contrasted this with an analysis using PCs calculated from all CpG sites, and found little difference in the results (not shown). The adjustment resulting from the PCs improved the validity of VC-score test results; however, (unusual) deviation of the PCEV test statistic from the null distribution was also indicated. This might be a consequence of the nonrobustness of PCEV to violation of normality assumption or the constant variance assumption of model residuals (errors). Even after normalization, methylation proportions have variances that are small when means are near 0 and 1. This heteroscedasticity might lead to a loss of power [9]. Thus, a transformation such as the logit may help in obtaining test statistics with valid null distributions.

We considered 2 ways to accommodate the relatedness among the participants: to restrict the analysis to unrelated individuals or to use a linear mixed model that takes family structure into account. In contrast to the well-behaved p value distribution for the analysis of unrelated subjects, the family-based VC-score test showed inflation in the Q-Q plots. Hence, the significant results for the 2 genes RNMT and MIR130B may be questionable. Almeida et al. [10] found that heritability of pretreatment DNA methylation was much higher than expected. Our results agree that pedigree-based kinship corrections are insufficient to correct for familial correlations in DNA methylation, and that additional corrections must be considered.

Other region-based association methods may be worth exploring in the future, such as the global analysis of methylation profiles (GAMP) [11] in which the density of methylation values in a region is approximated by B-splines and then the spline coefficients for each individual are used as covariates in association tests. Other strategies to accommodate the family structure in the region-based association tests include MF-KM (multivariate family data using kernel machine regression) [12], a linear mixed model built upon kernel machine regression, and mFARVAT (multivariate family-based rare variant association tool) [13], a quasi-likelihood-based-score test approach.

The region-based VC approach using unrelated individuals identified the gene LGALS9C as significantly associated with changes in triglycerides. However, univariate tests of individual CpG sites yielded no valid statistically significant results. After correctly accounting for the unknown confounding and subject relatedness, region-based methods show an improvement in power to detect associated genes as compared to single-marker methods.