Simulation of haplotype-resolved single-cell methylation profiles

DNA methylation is an epigenetic mark that regulates gene expression without altering the DNA sequence, and is crucial for tissue-specific gene expression, genomic imprinting, cellular differentiation, and X chromosome inactivation; its disruption is linked to many diseases. However, DNA methylation analysis remains blind to the diploid structure of the genome: no dataset specifies the true methylation state of every CpG on both haplotypes and both strands in each single cell for benchmarking tools. This thesis fills that gap by building an in silico multi–cell-type dataset and using it to quantify how much methylation signal survives realistic bisulfite sequencing. Single-cell and whole-genome bisulfite sequencing data are characterized across five genomic annotation classes (CpG island, shore, bimodal, promoter, gene body). Methylation is significantly overdispersed beyond sequencing noise, and a Beta distribution captures this overdispersion better than a Logit-Normal, although single-cell data pose challenges for parameter estimation. Therefore, spatial heterogeneity in WGBS data is adopted as a heuristic for within–cell-type cell-to-cell variability. A Beta-logit generalized linear mixed model is proposed to model overdispersion from WGBS estimates, representing cell-to-cell heterogeneity in single-cell data. The model is implemented on a diploid genome across four cell types, generating per-cell, per-haplotype, per-strand ground-truth methylation that includes allele-specific methylation at imprinted loci and hemimethylation. This ground truth undergoes in silico bisulfite conversion, paired-end read simulation, alignment, and methylation extraction at 0.5×, 2×, and 10× coverage. Analyses of ground-truth recovery after post-conversion processing show improved metrics with higher coverage, and the best performance is attributed to CpGs outside ASM regions. The simulated ground truth confirms cellular heterogeneity and shows that, at sufficient coverage, cells can still be clustered by cell type. These results provide a reusable haplotype-resolved single-cell methylation simulation framework.