Over the last decade, DNA methylation mapping played an important role in establishing the prevalence of altered DNA methylation in cancer40,41
. Furthermore, researchers have started to systematically study the role of DNA methylation in a wide range of non-neoplastic diseases42
. This is indeed a good time to probe for epigenetic alterations that contribute to human diseases: Genome-wide association studies have been completed for all common diseases, and their results point to a major role of non-genetic factors in the etiology of most diseases43
. Furthermore, it has been suggested that epigenetic events could provide a tractable link between the genome and the environment, with the epigenome emerging as a biochemical record of relevant life events44,45
. Systematic investigation of these topics requires powerful, accurate and cost-efficient methods for identifying DNA methylation differences between samples.
The goal of this study was to evaluate current methods for global DNA methylation mapping and to compare their performance in a practical application scenario. To mimic a typical disease-centered case-control study, we worked with primary patient material (colon samples) and used lower amounts of input DNA than in most previous studies (MeDIP: 300ng, MethylCap: 1µg, RRBS: 50ng, Infinium: 1µg). Furthermore, we focused on cell types that are known to exhibit relatively moderate DNA methylation differences30,46
, in contrast to the massive DNA methylation alterations that are frequently observed in cultured somatic cells10
and cancer cell lines47
. Finally, all four methods included in the current study are widely available and not excessively costly, such that there are few obstacles to using this technology comparison as a blue print for individual lab efforts as well as large-scale epigenomic case-control studies investigating the epigenetics of human diseases.
Overall, the data confirmed that all four methods provide accurate DNA methylation measurements and can be used to detect DMRs in clinical samples. In terms of accuracy, the bisulfite-based methods (RRBS, Infinium) performed slightly better than the enrichment-based methods and did not require any statistical correction of CpG bias. The genomic coverage was moderately higher for MethylCap than for MeDIP, RRBS coverage was by design focused on CpG-rich regions, and the Infinium assay covered a relatively small number of preselected genomic regions. Despite the striking differences in genomic coverage, a substantial fraction of DMRs detected by MeDIP or MethylCap were also identified by RRBS (and vice versa). This somewhat counter-intuitive observation can be explained by the role of region-specific read coverage for the ability to identify statistically significant DMRs: If a genomic region is CpG-poor and thus rarely sequenced by MeDIP or MethylCap, both methods have low statistical power to detect differential DNA methylation. In contrast, CpG-rich genomic regions tend to be more amenable to DMR detection by MeDIP and MethylCap and are also frequently covered by RRBS measurements. Finally, we observed that MethylCap was able to detect roughly twice as many DMRs as MeDIP at comparable sequencing depths, RRBS detected more DMRs than MeDIP but fewer DMRs than MethylCap, and the Infinium assay detected only 20% of the consensus DMRs identified by the sequencing-based methods. These differences could be reproduced in two independent pairwise comparisons, providing strong indication that they are robust across biological replicates and cannot be explained by random experimental variation. On the other hand, we used one specific protocol for each method, and it is quite possible that protocol variations (e.g., different antibody for MeDIP, different elution procedure for MethylCap, or different size selection for RRBS) would produce different results.
Our study also reinforces the importance of sequencing depth as a key parameter determining to power to detect differential methylation with any of the sequencing-based methods. To allow for a fair and practically relevant comparison, we sequenced approximately 30 to 40 million reads for each sample and method. However, it became evident that deeper sequencing would identify further DMRs, especially for MeDIP and MethylCap (Supplementary Figure 11
). For disease-centered studies it is therefore necessary to make an informed decision about how to distribute the available resources between sequencing few samples more deeply and sequencing more samples less deeply. Such a decision can be guided by statistical power calculations when some prior knowledge exists about the characteristics of expected DMRs (e.g., magnitude of difference, location in CpG-rich vs. CpG-poor genomic regions), or they can be dictated by practical considerations such as the number of available samples. MeDIP, MethylCap and RRBS as performed in this study seem to provide a practically useful compromise between breadth and depth of sequencing. In contrast, whole-genome bisulfite sequencing48
provides comprehensive genomic coverage at the cost of having to sequence over a billion reads per sample. On the other end of the spectrum, low sequencing depths are often sufficient to detect strong differences such as global loss of DNA methylation but fail to provide reliable locus-specific information49
Finally, genome-wide studies tend to ignore repetitive regions due to technical difficulties, and the few studies that focused specifically on mapping DNA methylation in repetitive regions did so at relatively low coverage50–52
. The current dataset was well-suited to analyze DNA methylation in repetitive regions because the joint results obtained by three different experimental methods helped us to control for technical artifacts that can burden the analysis of repetitive DNA. We observed that repeat sequences are most highly methylated when they are CpG-rich and highly prevalent in the human genome (Supplementary Data 2
). In contrast, the DNA methylation levels varied widely among repeat sequences that are either CpG-poor or infrequent in the genome. These results lend support to the hypothesis that DNA methylation provides a mechanism for keeping active retrotransposons in check53
. They argue for a highly specific mechanism of repeat repression, which targets DNA methylation mostly to those repeat sequences that threaten genome integrity, while many “benign” repeat sequences may remain unmethylated.
In summary, we benchmarked four methods for genome-scale DNA methylation profiling in terms of their accuracy and power to detect DNA methylation differences. These results will facilitate the selection of suitable methods for studying the role of DNA methylation in human diseases.