AbstractSingle cell RNA sequencing (scRNA-seq) provides a rich view into the heterogeneity underlying a cell population. However single-cell data are usually noisy and very sparse due to the presence of dropout genes. In this work we propose an approach to impute missing gene expressions in single cell data using generative adversarial networks (GANs). By learning an approximate distribution of the data, our approach, scGAIN, can impute dropouts in simulated and real single cell data. The work in this paper discusses how to adopt GAIN training model into the domain of imputing single cell data. Experiments show that scGAIN gives competitive results compared to the state-of-the-art approaches while showing superiority in various aspects in simulation and real data. Imputation by scGAIN successfully recovers the underlying clustering of different subpopulations, provides sharp estimates around true mean expressions and increase the correspondence with matched bulk RNAseq experiments.
Generative Adversarial Networks for Graph Data Imputation from Signed Observations
