Imputation in Scrna-seq Data Using Supervised Deep Generative Networks

Abstract

The main challenge of single-cell RNA sequencing (scRNA-seq) studies arises from the large data sizes and various technical noises such as the excess zero counts within individual cells (called dropout events). This inaccurate measurement of gene expressions may introduce bias in downstream analyses of scRNA-seq data, so it is necessary to correct the false zero expression by computational imputation methods. Most current imputation pipelines typically use unsupervised modeling approaches that expect to recover biologically meaningful gene expression without known cell labels. However, unsupervised imputation often yields only limited gains since very little cell identity information is retained in the observed expression data. And for datasets with known cell labels, the use of cell label-guided imputation is expected to recover more accurate gene expression dynamics in different cell populations. In this work, we developed a supervised deep generative imputation model called scIDG to recover the biologically meaningful gene expression in scRNA-seq data with known cell type information. scIDG can learn both intrinsic features of the observed gene expression data and the interpretable representations of cell types, generating biologically meaningful imputation values under specified cell types and recovering gene expression dynamics. We tested scIDG on simulated and real scRNA-seq datasets and compared it with several state-of-the-art methods for imputation using cell type information. Experiments showed that scIDG can more accurately recover the heterogeneity of different cell types, significantly improving downstream differential expression analysis and temporal trajectory inference analysis.