Fig. 2

Melissa robustly imputes CpG methylation states. a Imputation performance in terms of AUC as we vary the proportion of covered CpGs used for training. Higher values correspond to better imputation performance. For each CpG coverage setting, a total of 10 random splits of the data to training and test sets was performed. Each colored circle corresponds to a different simulation. The plot shows also the LOESS curve for each method as we increase CpG coverage. The methods considered were Melissa which shares information across cells and neighboring CpGs, the BPRMeth model that only shares information across neighboring CpGs, and a Random Forest classifier (RF) which predicts CpG methylation states using as input the observed CpG locations. Additionally, we considered three baseline models: Melissa Rate that transfers information across cells but not across neighboring CpGs using mean methylation levels across the genomic region, a Gaussian mixture model (GMM) that takes as input average M values across the region, and finally, the Rate method where we compute a mean methylation rate separately for each cell and genomic region. b Imputation performance measured by AUC for varying number of cells assayed. In a, N = 200 cells were simulated and cluster dissimilarity was set to 0.5, and in b, CpG coverage was set to 0.4 and cluster dissimilarity to 0.5