Exhaustive search for epistatic effects on the human methylome

Abstract

Studies assessing the existence and magnitude of epistatic effects on complex human traits provide inconclusive results. The study of such effects is complicated by considerable increase in computational burden, model complexity, and model uncertainty, which in concert decrease model stability. An additional source introducing significant uncertainty with regard to the detection of robust epistasis is the biological distance between the genetic variation and the trait under study. Here we studied CpG methylation, a genetically complex molecular trait that is particularly close to genomic variation, and performed an exhaustive search for two-locus epistatic effects on the CpG-methylation signal in two cohorts of healthy young subjects. We detected robust epistatic effects for a small number of CpGs (N = 404). Our results indicate that epistatic effects explain only a minor part of variation in DNA-CpG methylation. Interestingly, these CpGs were more likely to be associated with gene-expression of nearby genes, as also shown by their overrepresentation in DNase I hypersensitivity sites and underrepresentation in CpG islands. Finally, gene ontology analysis showed a significant enrichment of these CpGs in pathways related to HPV-infection and cancer.

Figure 1

Power analysis exhaustive search for epistatic effects. In (a) we adjusted alpha to reach genome-wide and methylome-wide Bonferroni correction (discovery phase, p = 6.8 × 10⁻¹⁸). In (b) we adjusted alpha to reach a per-CpG Bonferroni correction threshold (replication phase, p = 3.8 × 10⁻⁶). The legends depict the variance that can be explained (in percentage) for different effect sizes (r _min = 0.03, 0.1%; r _max = 0.55, 30%). The vertical gray bars correspond to a sample size of N = 533 (discovery sample) and N = 319 (replication sample).

Figure 2

Example of a main effect of a SNP causing a spurious significant interaction effect between two other SNPs. Data is shown for cg00022866 from the discovery sample. (a) rs11231741 shows a strong main effect (p = 4.5 × 10⁻¹¹²). This causes a spurious significant interaction (b p = 3.3 × 10⁻¹⁸) because rs11231741 is in LD with both interacting SNPs (rs11231740: r ² = 0.55; rs2236648: r ² = 0.25). Of note, the two interacting SNPs show low LD only (r ² = 0.024). Panel (c) depicts the dependencies between the 9 SNP-groups build from rs11231740 and rs2236648 and the 3 SNP-groups from rs11231741 (color-coded in black, red and green; a jitter has been added to the data): the 9 SNP-groups of the interacting SNPs mimic the three SNP-groups of the main effect, with 5 of the 9 groups mainly corresponding to the homozygous common allele carrier (black), 3 of the 9 groups mainly corresponding to the heterozygous group (red) and 1 group mainly corresponding to the homozygous rare allele carrier (green). Panel (d) shows the same data as in (b), but now with color-coding of the three SNP-groups from rs11231741.