To study inter-individual variation in methylation profiles we measured methylation levels across the genome in 77 lymphoblastoid cell lines (LCLs) derived from unrelated individuals from the HapMap Yoruba (YRI) collection. For these samples we also had publicly available genotypes 21 , as well as estimates of gene expression levels from RNA-sequencing in 69 of the 77 samples 24 . Methylation profiling was performed in duplicate using the Illumina HumanMethylation27 DNA Analysis BeadChip assay, which is based on genotyping of bisulfite-converted genomic DNA at individual CpG-sites to provide a quantitative measure of DNA methylation. The Illumina array includes probes that target 27,578 CpG-sites. However, we limited analyses to probes that mapped uniquely to the genome and did not contain known sequence variation, leaving us with a data set of 22,290 CpG-sites in the promoter regions of 13,236 genes (see Methods). Following hybridization, methylation levels were estimated as the ratio of intensity signal obtained from the methylated allele over the sum of methylated and unmethylated allele intensity signals. Methylation levels were quantile-normalized 28 across two replicates. We tested for correlations with potential confounding variables that could affect methylation levels in LCLs 29 , such as LCL cell growth rate, copy numbers of Epstein-Barr virus, and other measures of biological variation (see Additional file 1) that were available for 60 of the individuals in our study 30 ; these did not significantly explain variation in the methylation levels in our sample (Figure S1 in Additional file 1). However, we observed an influence of HapMap Phase (samples from Phase 1/2 vs 3) on the distribution of the first principal component loadings in the autosomal data, suggesting that the first methylation principal component may in part capture technical variation potentially related to LCL culture. In the downstream association mapping analyses, we applied a correction using principal component analysis regressing the first three principal components to account for unmeasured confounders and increase power to detect quantitative trait loci.

Distinct patterns of methylation were observed for CpG-sites located on the autosomes, X-chromosome, and in the vicinity of imprinted genes (Figure 1a). The majority (71.4%) of autosomal CpG-sites were primarily unmethylated (observed fraction of methylation <0.3), 15.6% were hemi-methylated (fraction of methylation was between 0.3 and 0.7), and 13% were methylated. As expected, these patterns were consistent with previously observed lower levels of methylation near promoters relative to genome-wide levels 4 31 . We did not find evidence for sex-specific autosomal methylation patterns, consistent with a previous report 4 . In contrast, CpG-sites on the X-chromosome exhibited highly significant sex-specific differences (Figure S2) with hemi-methylated patterns in females that were consistent with X-chromosome inactivation. A similar hemi-methylation peak was observed for CpG-sites located near the transcription start sites (TSSs) of known autosomal imprinted genes in the entire sample.

We observed a previously reported 4 drop in methylation levels for CpG-sites located within 1 kb of TSSs (Figure 1b). Promoter methylation levels have been reported to vary with respect to CpG islands 32 . We found that although distance to the CpG island (CGI) border 33 (including CpG shores 34 ) did not significantly affect methylation levels, CpG-sites located in CGIs were under-methylated and less variable (Wilcoxon rank-sum test P < 2.2 × 10^-16) compared to sites outside of CGIs (Figure 1, Figure S3 in Additional file 1).

Methylation is often found to be correlated across genomic regions at the scale of 1-2 kb 4 35 . We investigated whether the correlation between autosomal methylation levels (co-methylation) depended on the distance between CpG-sites. We observed that methylation levels at probes located in close proximity (up to 2 kb apart) were highly correlated (Figure 1c), indicating that variation in methylation levels between individuals is correlated within cell type. Figure 1c also shows that pairs of CpG-sites that were both within a CGI showed greater evidence for co-methylation than pairs of CpG sites for which at least one was outside the CGI, controlling for distance, implying differential regulation of DNA methylation for CpGs inside and outside of CGIs 32 .

Methylation has long been implicated in the regulation of gene expression. To examine the role of methylation in gene expression variation, we compared methylation levels to estimates of gene expression based on RNA-sequencing (Figure 2a). Within individuals, we found a significant negative correlation between methylation and gene expression levels (Figure S4 in Additional file 1) across 11,657 genes (mean rank correlation r = -0.454). We divided the genes into quartiles from high to low gene expression and observed that the drop in methylation levels near to the TSS (Figure 1b) was only seen in highly expressed genes (Figure 2b). We also asked whether variation in methylation levels across individuals correlates with variation in gene expression levels. Comparisons at the gene level across 69 individuals indicated a modest but significant excess of negatively correlated genes (permutation P < 0.0001).

DNA methylation is thought to interact with histone modifications during the regulation of gene-expression 36 37 . We compared methylation levels in our sample with histone modification ChIP-seq data from the ENCODE project in one of the CEPH HapMap LCLs (GM12878). We found strong negative correlations between DNA methylation levels and the presence of histone marks that target active genes (Figure 1d; Figures S3 and S5 in Additional file 1). For example, DNA methylation was low in H3K27ac peaks, which are indicative of enhancers 38 , have previously been positively correlated with transcription levels 39 and negatively correlated with DNA methylation levels 31 . Similarly, the transcription marks H3K4me3 and H3K9ac were both negatively correlated with DNA methylation levels. We also observed lower methylation levels in transcription factor binding sites predicted by the CENTIPEDE algorithm, using cell-type specific data including DNase1 sequencing reads 40 , consistent with the expectation that the absence of methylation is important for transcription factor binding.

We next assessed whether genetic variation contributes to inter-individual variation in DNA methylation levels. We first tested whether any SNPs were associated with overall patterns of DNA methylation, as measured by principal component analysis (see Methods). The most interesting signal was obtained for SNP rs10876043, which had a genome-wide significant association with variation in the first principal component of methylation (P = 4.5 × 10^-9), and which also showed a modest association with average genome-wide methylation levels (P = 4.0 × 10^-5) (Table S1 in Additional file 1). This SNP lies within the intron of the gene DIP2B, which contains a DMAP1-binding domain, and has been previously proposed to play a role in DNA methylation 41 .

Finally, we examined the overlap in regulatory variation that affects both methylation and gene expression levels using RNA-sequencing data 24 . We hypothesized that since DNA methylation can regulate gene expression, then variants that affect methylation should often have consequent effects on gene expression. The first way that we looked at this was to take the set of 180 SNPs that are meQTLs at FDR <10% (taking only the most significant SNP for each meQTL). We then tested each of these SNPs for association with expression levels of nearby genes (Figure 4a, red points). There is a clear enrichment of association with expression levels compared to the null hypothesis (black line) and compared to sets of control SNPs that are matched in terms of allele frequency and distance-to-probe distributions (black dots).

One example of a SNP, rs8133082, that is both a meQTL and eQTL for the gene C21orf56 is illustrated in Figure 5. When we regress out methylation, this completely removes the association of this SNP with gene expression (Figure 5a, b, c, d). We validated the methylation assay findings at C21orf56 by bisulfite sequencing the methylation probe region in eight samples in our study, four from each homozygote genotype class for the SNP (Figure 5f). The two methylation probes at C21orf56 both had cis meQTLs and overlapped the likely promoter region as indicated by histone modification data (Figure 5e), suggesting that genetic variation may affect the chromatin structure in this region. C21orf56 appears to modulate the response of human LCLs to alkylating agents, and may act as a genomic predictor for inter-individual differences in response to DNA damaging agents 46 .

To examine further the overlap between eQTLs and meQTLs, we re-analyzed the eQTL data by incorporating methylation as a gene-specific covariate. If variation in methylation underlies variation in gene-expression, we expect to observe a drop in the number of eQTLs in the methylation-residual gene expression data. At an FDR of 10% (P = 2.5 × 10^-5) there were 484 original eQTLs and 463 methylation-residual eQTLs, where 439 eQTLs overlapped, 45 eQTLs were present only in the original data, and 24 new eQTLs were present only in the methylation-residuals (Figure 4b). Interestingly, the SNPs that were eQTLs for the 45 genes with reduced signals in the methylation-residuals were enriched for significant methylation associations (Figure S10 in Additional file 1), suggesting that these are true underlying meQTLs, where genetic variation affects methylation, which in turn regulates gene expression 5 18 . In summary our results indicate a significant enrichment of SNPs that affect both methylation and gene expression, suggesting a shared mechanism (for example, that increased DNA methylation might drive lower gene expression). However the number of genes that show such a signal is a modest fraction of the total number of meQTLs.

DNA was extracted from lymphoblastoid cell lines from 77 individuals from the Yoruba (YRI) population from the International HapMap project (60 HapMap Phase 1/2 and 17 HapMap Phase 3 individuals). Lymphoblastoid cell lines were previously established by Epstein-Barr Virus transformation of peripheral blood mononuclear cells using phytohemagluttinin. We obtained the transformed cell lines from the Coriell Cell Repositories. Methylation data were obtained using the Illumina HumanMethylation27 DNA Analysis BeadChip assay. Methylation estimates were assayed using two technical replicates per individual and methylation levels were quantile normalized across replicates 28 . At each CpG-site the methylation level is presented as β, which is the fraction of signal obtained from the methylated beads over the sum of methylated and unmethylated bead signals. We considered different approaches to normalizing values across replicates, as well as using the log of the ratio of methylated to unmethylated signal instead of β, and found the results robust to normalization procedure, measure of methylation, and across technical replicates (see Additional file 1). The methylation data are publicly available 50 and have been submitted to the NCBI Gene Expression Omnibus 51 under accession no. [GSE26133].

We mapped the 27,578 Illumina probes to the human genome sequence (hg18) using BLAT 52 and MAQ 53 . We selected 26,690 probes that unambiguously mapped to single locations in the human genome at a sequence identity of 100%, discarding probes that mapped to multiple locations with up to two mismatches. We excluded a further 4,400 probes that contained sequence variants, including 3,960 probes with SNPs (from the 1,000 genomes project 54 , July 2009 release, YRI population) and 440 probes which overlapped copy number variants 55 . This resulted in a final set of 22,290 probes (21,289 autosomal probes) that were used in all further analyses. The 22,290 probes were nearest to the TSSs of 13,236 Ensembl genes, of which 12,901 genes had at least one methylation CpG-site within 2 kb of the TSS.

Bisulfite sequencing was performed in the C21orf56 region for eight individuals. DNA was bisulfite-converted using the EZ DNA Methylation-Gold Kit (Zymo Research). PCR amplification was performed using primers designed around CpG-site cg07747299 from the HumanMethylation27 array and the nearest CpG island in the region (using Methyl Primer Express from Applied Biosystems) for a total of 411 bp amplified in the 5' UTR of the C21orf56 gene. PCR products were sequenced and cytosine peak heights compared to overall peak height were called using 4Peaks Software.

RNA-sequencing data were obtained for LCLs from 69 individuals in our study from 24 . The methylation and RNA-sequencing data were obtained from the same cultures of the LCLs. RNA-sequencing gene expression values are presented as the number of GC-corrected reads mapping to a gene in an individual, divided by the length of the gene. In the methylation to gene expression comparisons we split genes into quantiles based on the mean gene expression per gene. For the eQTL analyses, RNA-sequencing data were corrected and normalized exactly as in 24 . Of the 22,683 genes in the original study, 10,167 autosomal genes had both gene expression counts and methylation CpG-sites within 2 kb of the TSS.

HapMap release 27 genotype data were obtained for 3.8 million autosomal SNPs in HapMap (combined Phase 1/2 and 3). Missing genotypes were imputed by BIMBAM 56 using the posterior mean genotype. Non-polymorphic SNPs were excluded, reducing the set to 3,035,566 autosomal SNPs for association analyses.

Spearman rank correlations were used to assess co-methylation between probes and to compare methylation and gene expression. We used 10,000 permutations of the gene expression to methylation assignments to assess the enrichment of negatively and positively correlated genes in the 25% and 5% tails within genes. Wilcoxon rank-sum tests were used to compare probe means and variances for subsets of probes.

Genome-wide association was performed using the methylation values at each CpG-site as phenotypes and three million autosomal SNP genotypes. We used least squares linear regression with a single-locus additive effects model, where we estimated the effect of the minor SNP allele on the increase in methylation levels. Prior to the association analyses, we normalized the methylation values at each CpG-site to N(0, 1) and applied a correction using principal component analysis regressing the first three principal components to account for unmeasured confounders following similar approaches to reduce expression heterogeneity in gene expression experiments 24 42 43 (see Additional file 1). Sex-specific analyses were performed using sex as a covariate and assessing the significance of the sex by additive-QTL interaction term.

We assessed the enrichment of association at SNPs and probes that were previously reported to be associated with methylation 7 8 15 16 17 18 and at SNPs within 200 kb of genes known to affect DNA methylation (Table S3 in Additional file 1). We also compared genetic variation to normalized variation in the principal components loadings for the autosomal methylation data (see Additional file 1). Results from the 180 cis meQTLs are available online 50 .

We performed genome-wide permutations to assess the significance of the genome-wide association results in the least-square linear regressions. We permuted the methylation values for the 21,289 autosomal probes (phenotypes), performed genome-wide association on the 21,289 permuted and normalized phenotypes, and repeated this procedure for 10 (cis-analyses) or 1 (trans-analyses) replicates selecting the best signal per probe per replicate. Results are presented at an FDR of 10%, meaning that an estimated 10% of the meQTLs are false positives. Results for additional FDR thresholds are shown in Additional file 1. FDR was calculated as the fraction of significant hits in the permuted versus the observed data at a given P-value threshold. The association analyses and FDR calculations were performed for all autosomal principal components and CpG-sites in the methylation data, and for all autosomal genes in the RNA-sequencing data.

We fitted a Bayesian hierarchical model 22 to test whether meQTLs were over-represented in transcription factor binding sites, histone-modifications, and with respect to distance to the probe. We extended the model to fit the methylation data, where the reference point was the location of the methylation probe. Each annotation category that we examined was included in the model while accounting for distance effects.

Genome annotation data were obtained from UCSC (hg18). Histone modification data were obtained from ChIP-seq reads from the ENCODE project (Bernstein lab) for GM12878 for seven histone marks. Histone modification categories were based on estimated peaks in the read-depth distribution (see Additional file 1).

Transcription factor binding site locations were estimated using the algorithm CENTIPEDE 40 57 . For the results presented here, CENTIPEDE started by identifying all matches in the genome to a large number of transcription factor binding motifs obtained from the TRANSFAC and JASPAR databases. It then estimated which potential binding sites are actually occupied by transcription factors in LCLs, by incorporating input data from sequence conservation, location with respect to nearby genes, and cell-specific experimental data, including DNaseI data. We used 1,136,620 non-overlapping sites from 751 transcription factor motif matches that overlapped 1,913 CpG-sites.