Dropout-based feature selection for scRNASeq

Abstract

Features selection is a key step in many single-cell RNASeq (scRNASeq) analyses. Feature selection is intended to preserve biologically relevant information while removing genes only subject to technical noise. As it is frequently performed prior to dimensionality reduction, clustering and pseudotime analyses, feature selection can have a major impact on the results. Several different approaches have been proposed for unsupervised feature selection from unprocessed single-cell expression matrices, most based upon identifying highly variable genes in the dataset. We present two methods which take advantage of the prevalence of zeros (dropouts) in scRNASeq data to identify features. We show that dropout-based feature selection outperforms variance-based feature selection for multiple applications of single-cell RNASeq.