scMelody: An Enhanced Consensus-Based Clustering Model for Single-Cell Methylation Data by Reconstructing Cell-to-Cell Similarity

Abstract

Single-cell DNA methylation sequencing technology has brought new perspectives to investigate epigenetic heterogeneity, supporting a need for computational methods to cluster cells based on single-cell methylation profiles. Although several methods have been developed, most of them cluster cells based on single (dis)similarity measures, failing to capture complete cell heterogeneity and resulting in locally optimal solutions. Here, we present scMelody, which utilizes an enhanced consensus-based clustering model to reconstruct cell-to-cell methylation similarity patterns and identifies cell subpopulations with the leveraged information from multiple basic similarity measures. Besides, benefitted from the reconstructed cell-to-cell similarity measure, scMelody could conveniently leverage the clustering validation criteria to determine the optimal number of clusters. Assessments on distinct real datasets showed that scMelody accurately recapitulated methylation subpopulations and outperformed existing methods in terms of both cluster partitions and the number of clusters. Moreover, when benchmarking the clustering stability of scMelody on a variety of synthetic datasets, it achieved significant clustering performance gains over existing methods and robustly maintained its clustering accuracy over a wide range of number of cells, number of clusters and CpG dropout proportions. Finally, the real case studies demonstrated the capability of scMelody to assess known cell types and uncover novel cell clusters.