Discovery and optimization of cell-type-specific DNA methylation markers for in silico deconvolution (2304.05756v1)

Abstract: DNA methylation is a significant driver of cell-type heterogeneity and has been implicated in various regulatory processes ranging from cell differentiation to imprinting. As the methyl group is embedded in the DNA molecule, assessing DNA methylation is particularly promising in liquid-biopsy-based approaches, as cell-free DNA retains information related to its cell of origin. In this work, I leverage a recently profiled collection of cell-sorted whole genome bisulfite profiles of 44 healthy cell types. The high quality and purity of such data provide an ideal basis for discovering and characterizing discriminative DNA methylation regions that could serve as a reference for computational deconvolution. First, I characterize differentially methylated regions between every pair of cell types, obtaining a meaningful measure of divergence. Pairwise differences were then aggregated to identify a set of uniquely (de)methylated regions (UMRs) for each cell type. Identified UMRs are predominantly hypomethylated and their numbers vary significantly across cell types. They are mostly located in enhancer regions and strongly support cell-type-specific characteristics. As mapping onto UMRs has proven unsuitable for deconvolution, I developed a novel approach utilizing the set cover algorithm to select discriminative regions for this purpose. Based on these regions, deconvolution was performed in two distinct approaches: a beta-value-based and a read-level one. Both approaches outperform an existing deconvolution software modeled on the same data 3-fold in terms of total deconvolution error. Surprisingly, the beta-based approach slightly outperformed the read-level one. Overall, I present an adaptable, end-to-end software solution (methylcover) for obtaining accurate cell type deconvolution, with possible future applications to non-invasive assays for disease detection and monitoring.