Deep zero-inflated negative binomial model and its application in scRNA-seq data integration

Abstract

In the past decades, many single-cell RNA sequencing (scRNA-seq) datasets have been generated with wide applications such as cell type identification and differential gene expression analysis. In practice, it is important to integrate different scRNA-seq datasets, which is beneficial for the efficiency of downstream analysis. However, this integration task is challenging due to the unwanted technical effects within the integrated dataset that would blur the biological effect. To address this issue, we propose a novel deep learning based statistical model, using the Maximum Mean Discrepancy (MMD) loss and zero-inflated negative binomial (ZINB) regression model, to remove two technical effects, i.e., batch effect and drop out effect, which are commonly existed unwanted technical noises confounding with the biological signals. We demonstrate that our model can effectively remove the technical effects while preserve the significant biological effects through a real scRNA-seq data analysis.