1 Installation

## try http:// if https:// URLs are not supported
if (!requireNamespace("BiocManager", quietly=TRUE))
    install.packages("BiocManager")
BiocManager::install("Melissa")

## Or download from Github repository
# install.packages("devtools")
devtools::install_github("andreaskapou/Melissa", build_vignettes = TRUE)

2 Background

Measurements of DNA methylation at the single cell level are promising to revolutionise our understanding of epigenetic control of gene expression. Yet, intrinsic limitations of the technology result in very sparse coverage of CpG sites (around 5% to 20% coverage), effectively limiting the analysis repertoire to a semi-quantitative level. Melissa (MEthyLation Inference for Single cell Analysis) [1], is a Bayesian hierarchical method to quantify spatially-varying methylation profiles across genomic regions from single-cell bisulfite sequencing data (scBS-seq).

Melissa addresses the data sparsity issue by leveraging local correlations between neighbouring CpGs and similarity between individual cells (see Fig.1). The starting point is the definition of a set of genomic regions (e.g. genes or enhancers). Within each region, Melissa postulates a latent profile of methylation, a function mapping each CpG within the region to a number in \([0,1]\) which defines the probability of that CpG being methylated. To ensure spatial smoothness of the profile, Melissa uses a generalised linear model of basis function regression along the lines of [2,3] (with modified likelihood to account for single cell data). Local correlations are however often insufficient for regions with extremely sparse coverage, and these are quite common in scBS-seq data. Therefore, we share information across different cells by coupling the local GLM regressions through a shared prior distribution. In order to respect the (generally unknown) population structure that may be present within the cells assayed, we choose a (finite) Dirichlet mixture model prior.