MD: Statistical Testing for RRBS Data Sets

RRBS experiments offer a cost-effective method for measuring the methylation levels of cytosines in CpG dense regions. Statistical testing for differences in CpG methylation between sample groups typically involves beta-binomial modelling of CpG loci individually and testing across samples, such as with the BiSeq and MethylSig packages. Individual CpGs are then chained together to output a list of putative differentially methylated regions (DMRs). We take a different approach, instead testing pre-defined regions of interest for changes in the distribution of the methylation profiles. Changes are compared to inter-replicate differences to establish which regions vary in a manner that cannot be explained by replicate variation alone. We utilise the maximum mean discrepancy (MMD) [1] to perform the test, adjusting this measure to account for changes in the coverage profile between the testing groups. In this vignette, we run through an analysis of a toy data set using the MD method. We use the same data structures as the ’BiSeq’ package, and assume that we have data stored in an rrbs structure with an accompanying GRanges object describing the regions of interest we want to test. For a fuller explanation and exploration of the method, please see the open access journal article [2].