Fig 1.
(A) The input for Metheor is bisulfite read alignment tagged with Bismark methylation call strings. Using each of the seven subcommands shown, Metheor computes the corresponding DNA methylation heterogeneity measure. If reads were aligned with a tool other than Bismark, Metheor can still add tag for methylation call string with metheor tag subcommand to make alignment file compatible for Metheor run. (B) Schematic diagram for DNA methylation heterogeneity measures and benchmark settings in this study. [5] denote the Perl script provided by the authors along with the article proposing the utility of MHL. (C, D) Schematic diagram illustrating (C) read-centric algorithm and (D) CpG-centric algorithm for the computation of DNA methylation heterogeneity. The advantages (plus symbol) and disadvantages (minus symbol) are shown below the diagrams. (E) Distribution of the average number of CpGs per sequencing read for the RRBS data from 928 CCLE cell lines. (F) Genomewide average levels of proportion of discordant reads (PDR) and local pairwise methylation discordance (LPMD) against varying read lengths. (G) Schematic illustration for the definition of local pairwise methylation discordance (LPMD) and examples. The proportion of reads having different DNA methylation states for a pair of CpGs (red arrows) are computed.
Fig 2.
Performance benchmark and validity of the results.
Benchmarking the running time of Metheor using (A) simulated RRBS dataset and (B) Ewing sarcoma RRBS dataset. Values below the name of each of the measures denote the amount of speedup (in fold) in Metheor compared to its benchmark counterpart. Benchmarking the memory usage of Metheor using (C) simulated RRBS dataset and (D) Ewing sarcoma RRBS dataset. Values below the name of each of the measures denote the amount of memory usage reduction (in fold) in Metheor compared to its benchmark counterpart. All the benchmark experiments were repeated for three times, except for MHL. Lines denote the average wall time and shades represent the 95% confidence interval. The wall time for MHL computation was measured for only once. (E) Validity of the results. CpG-wise (PDR, MHL, FDRP and qFDRP) and CpG quartet-wise (PM and ME) methylation heterogeneity levels were compared between Metheor and the corresponding reference implementations. Pearson’s correlation coefficient and corresponding p-values are shown for FDRP and qFDRP.
Table 1.
Performance comparison with WSHPackage using multiple threads for 20M RRBS-simulated reads.
Fig 3.
Characteristics of LPMD across 928 cancer cell lines.
(A) Genomewide average methylation levels and LPMD levels grouped by tissue types. Black vertical lines denote groupwise average levels of methylation and LPMD levels. Black horizontal bars on the right side denote the standard deviation of corresponding values. (B) Genomewide average methylation levels and LPMD levels grouped by disease types. Disease types from haematopoietic and lymphoid tissues are highlighted in red. (C, D) Correlation between mRNA expression and (C) genomewide average LPMD or (D) genomewide average methylation level. Genes are ranked according to the p-values of the corresponding correlation coefficients. P-values were adjusted using Benjamini-Hochberg procedure. (E, F) Correlation between DNMT3A expression and (E) genomewide average LPMD or (F) genomewide average methylation level. (G, H) Trends of fixed-distance average LPMD values. Shades denote 95% confidence interval. In (H), Cell lines were divided into two groups based on the median DNMT3A expression. (I) Difference of fixed-distance average LPMD values between DNMT3AHigh and DNMT3ALow groups.
Fig 4.
Association between methylation entropy and cancer stemness.
(A) Genes were ranked by the Pearson’s correlation between their expression and average methylation entropy levels across promoters. Red dots represent 3,680 genes having statistically significant correlations (Benjamini-Hochberg adjusted p-value < 0.05), and the results of functional enrichment analysis using those genes are shown in (B). (C) The distribution of promoter methylation entropy levels in primary and metastatic cancer cell lines. (D) The association between promoter methylation entropy levels and two genes representative of Wnt signaling pathway (WNT7A and CTNND2). (E) The association between promoter methylation entropy levels and the activity of Wnt signaling pathway. *two-tailed independent t-test p < 0.05; In D-E, Pearson’s correlation coefficients and associated p-values are shown. In D, p-values were adjusted using Benjamini-Hochberg procedure.