Abstract
Single-cell RNA sequencing (scRNA-seq) reveals the transcriptome diversity in heterogeneous cell populations as it allows researchers to study gene expression at single-cell resolution. The latest advances in scRNA-seq technology have made it possible to profile tens of thousands of individual cells simultaneously. However, the technology also increases the number of missing values, i. e, dropouts, from technical constraints, such as amplification failure during the reverse transcription step. The resulting sparsity of scRNA-seq count data can be very high, with greater than 90% of data entries being zeros, which becomes an obstacle for clustering cell types. Current imputation methods are not robust in the case of high sparsity. In this study, we develop a Neural Network-based Imputation for scRNA-seq count data, NISC. It uses autoencoder, coupled with a weighted loss function and regularization, to correct the dropouts in scRNA-seq count data. A systematic evaluation shows that NISC is an effective imputation approach for handling sparse scRNA-seq count data, and its performance surpasses existing imputation methods in cell type identification.