Defiant: (DMRs: easy, fast, identification and ANnoTation) identifies differentially Methylated regions from iron-deficient rat hippocampus

Abstract

Background: Identification of differentially methylated regions (DMRs) is the initial step towards the study of DNA methylation-mediated gene regulation. Previous approaches to call DMRs suffer from false prediction, use extreme resources, and/or require library installation and input conversion.

Results: We developed a new approach called Defiant to identify DMRs. Employing Weighted Welch Expansion (WWE), Defiant showed superior performance to other predictors in the series of benchmarking tests on artificial and real data. Defiant was subsequently used to investigate DNA methylation changes in iron-deficient rat hippocampus. Defiant identified DMRs close to genes associated with neuronal development and plasticity, which were not identified by its competitor. Importantly, Defiant runs between 5 to 479 times faster than currently available software packages. Also, Defiant accepts 10 different input formats widely used for DNA methylation data.

Conclusions: Defiant effectively identifies DMRs for whole-genome bisulfite sequencing (WGBS), reduced-representation bisulfite sequencing (RRBS), Tet-assisted bisulfite sequencing (TAB-seq), and HpaII tiny fragment enrichment by ligation-mediated PCR-tag (HELP) assays.

Keywords: Bisulfite sequencing; DNA Methylation; Differentially Methylated regions (DMR); Epigenetics; RRBS; WGBS.

Fig. 1

Illustration of how different DMR results are compared. Artificial (or “Benchmark”) DMRs are depicted in green, while predicted DMRs are shown in red

Fig. 2

Defiant uses WWE for DMR identification. We used artificial data designed for Metilene [27]. Each group is composed of ten replicate DNA methylation samples. The top two panels show the level of DNA methylation for each CpG (box-and-whisker plots). The mean is weighted based on coverage. The third panel from the top shows the weighted Welch p-value between the sets for individual CpG. The bottom panel shows differences between the weighted mean. Defiant calls a DMR when it finds consecutive CpGs with 1) differences in methylation levels, 2) minimum coverage, 3) p-value. When Defiant finds consecutive CpGs that do not match the above criteria, the expansion of a DMR stops. The third and fourth panels show the DMR start and end points in red

Fig. 3

Performance comparison using all 16 Metilene’s benchmark datasets composed of RRBS and WGBS. Each point represents a data set. Precision is compared against FN (a) and recall against precision (b). In plot a, 15 points from Defiant, 15 from Metilene, 10 from RADMeth, and 5 from RnBeads overlap in the upper-left corner. Both Defiant and Metilene outperformed other predictors

Fig. 4

F1 values of 16 different artificial data sets. Circles indicate mean values. Both Defiant and Metilene outperformed other predictors

Fig. 5

Comparison of DMR overlay of program with respect to benchmark in panel (a), and benchmark with respect to program in panel (b) for the combined RRBS and WGBS data sets

Fig. 6

An iron-deficient diet induces strong changes in methylation for many genes, the strongest of which are shown here. a Gucy2c displays strongly increased methylation from an Fe-deficient diet. b Usp36 is hypermethylated in Fe-deficient pups. c Tnni1 shows moderate, but consistently significant hypermethylation. d Fkrp shows strong and consistent hypomethylation

Fig. 7

Statistics on CpG regions predicted uniquely by Metilene. While, these genes show differences in the methylation levels, their p-values were unacceptably high when coverage information is considered. a The differences in the methylation levels over weighted Welch’s p-values. b The mean coverage against the weighted p-value c distribution of coverage. Real WGBS shows much wider distribution of coverage than the artificial data set. “C” = iron sufficient rat samples, “Fe” = Iron-deficient rat samples, “R” = artificial RRBS data, and “W” = artificial WGBS data

Fig. 8

The DMRs identified by Defiant but missed by Metilene. DMRs on a. Pde6c gene body b. Cdh2 gene body, c. Mobp gene gody, and d. Pck1 promoter region

Fig. 9

Resources used by DMR callers in time used (panel a) and RAM (panel b). We used eight artificial WGBS data sets. All evaluation was done on a computer with Intel(R) Xeon(R) CPU E5–2620 v2 @ 2.10GHz with 32 GB RAM