Similar to the LD relationships that exist between proximal genetic variants, DNAm levels are often correlated between proximally located DNAm sites.14, 46 To further characterize the genetic architecture of DNA methylation, we investigated whether shared genetic effects on multiple DNAm sites underlies this regional correlation structure. Although genetic variants are often associated with variation at multiple DNAm sites (Figure S4), this does not establish a shared genetic effect; shared genetic signals influencing a pair of DNAm sites might result from two distinct causal genetic variants that are in strong LD. To formally test whether neighboring DNAm sites are influenced by the same causal variant, we used a Bayesian co-localization approach44 to interrogate all pairs of DNAm sites characterized as being located within 250 kb of each other and associated with at least one significant mQTL variant at our “discovery” significance threshold (p < 1 × 10−10). Our analyses assessed 3,535,812 pairs of DNAm sites with a median distance between DNAm sites of 110,493 bp (interquartile range = 47,914–178,085) and compared the pattern of mQTL associations for both DNAm sites to test whether they index an association with either the same causal variant or two distinct causal variants. We found that the posterior probabilities for virtually all of these (n = 3,520,781 [99.6%], median distance of 110,319 bp [interquartile range = 47,803–177,948]) supported a co-localized association within the same genomic region (PP3 + PP4 > 0.99). Of these, 281,898 pairs (8%) had sufficient support for the association of both DNAm sites with the same causal mQTL variant (PP3 + PP4 > 0.99 and PP4/PP3 > 1; Table S4); 234,460 pairs (6.6%) had “convincing” evidence (PP3 + PP4 > 0.99 and PP4/PP3 > 5) for co-localization of the same mQTL association according to the criteria of Guo and colleagues.47 DNAm sites that shared genetic effects with at least one other DNAm site co-localize with a median of three other DNAm sites, indicating a complex relationship between genetic variation and DNAm in cis. Figure 2, for example, demonstrates that chromosome 9 contains a broad genomic region (>400 kb) where 38 DNAm sites—spanning seven genes—have a common underlying genetic signal. Of note, these DNAm sites are not contiguous; a small number of genetically mediated DNAm sites located within this region do not share the same mQTL signal. Pairs of DNAm sites with a shared causal mQTL variant are enriched for concordant directions of effect (71.2% pairs with positive correlations versus 28.8% pairs with negative correlations, binomial test p = 1.48 × 10−323; Figure S11). Furthermore, these pairs are located relatively close together (median distance between convincing co-localized pairs = 12,394 bp [interquartile range = 1,004–49,110]), with evidence that the shared genetic architecture is structured around annotated genomic features. Co-localized pairs of DNAm sites are significantly more likely to be annotated to the same gene (OR = 6.08, Fisher’s test p < 2.23 × 10−308) or CpG island (OR = 1.54, Fisher’s test p < 2.23 × 10−308) than non-co-localized pairs. Where pairs of DNAm sites with a shared genetic signal are annotated to the same gene, they are nominally less likely to be annotated to the same feature than are pairs of DNAm sites annotated to different genes (OR = 0.956, Fisher’s test p = 2.52 × 10−7), suggesting that where genetic variation influences DNAm at multiple sites across a gene these sites do not necessarily cluster by genic feature and can be located anywhere from the transcription start site to the end of the last exon. DNAm is more likely to be positively correlated between pairs of co-localized sites annotated to the same gene than between pairs of sites annotated to different genes (OR = 1.85, Fisher’s p < 2.23 × 10−308), a result driven predominantly by pairs of DNAm sites annotated to the same feature within that gene (OR = 1.57, Fisher’s test p = 3.41 × 10−135) rather than those annotated to different features within a gene. Finally, pairs of DNAm sites with shared genetic effects annotated to the same genic feature, although not necessarily the same gene, are more likely to be positively correlated than pairs annotated to different genic features (OR = 1.73, Fisher’s p < 2.23 × 10−308; Figure S12).
Figure 2 Shared Genetic Architecture between Neighboring DNA-Methylation Sites
Heatmap of Bayesian co-localization results for all pairs of DNA-methylation sites with at least one significant mQTL (p < 1 × 10−10) in a genomic region on chromosome 9 (chr9:124783559–125216341). Columns and rows represent individual DNA-methylation sites (ordered by genomic location). The color of each square indicates the strength of the evidence for a shared genetic signal (from yellow [weak] to red [strong]); this strength is calculated as the ratio of the posterior probabilities that they share the same causal variant (PP4) compared to two distinct causal variants (PP3). The ratio was bounded to a maximum value of 10; gray indicates pairs of DNA-methylation sites that were not tested for co-localization.