Abstract
Single-cell RNA sequencing (scRNA-seq) offers opportunities to study gene expression of tens of thousands of single cells simultaneously, to investigate cell-to-cell variation, and to reconstruct cell-type-specific gene regulatory networks. Recovering dropout events in a sparse gene expression matrix for scRNA-seq data is a long-standing matrix completion problem. In this article, we introduce Bfimpute, a Bayesian factorization imputation algorithm that reconstructs two latent gene and cell matrices to impute the final gene expression matrix within each cell group, with or without the aid of cell type labels or bulk data. Bfimpute achieves better accuracy than ten other publicly notable scRNA-seq imputation methods on simulated and real scRNA-seq data, as measured by several different evaluation metrics. Bfimpute can also flexibly integrate any gene- or cell-related information that users provide to increase performance.