CellBender remove-background: a deep generative model for unsupervised removal of background noise from scRNA-seq datasets

Abstract

Droplet-based single-cell assays, including scRNA-seq, snRNA-seq, and CITE-seq, produce a significant amount of background noise counts, the hallmark of which is non-zero counts in cell-free droplets and off-target gene expression in unexpected cell types. The presence of such systematic background noise is a potential source of batch effect and spurious differential gene expression. Here we develop a deep generative model for noise-contaminated data that is structured to reflect the phenomenology of background noise generation in droplet-based single-cell assays. The proposed model successfully distinguishes cell-containing from cell-free droplets without supervision, learns the profile of background noise, and retrieves a noise-free quantification in an end-to-end fashion. We present a scalable and robust implementation of our method as a module in the open-source software package CellBender. We show that CellBender operates close to the theoretically optimal denoising limit in simulated datasets, and present extensive evaluations using real datasets and experimental benchmarks drawn from different tissues, protocols, and modalities to show that CellBender significantly improves the agreement of droplet-based single-cell data with established gene expression patterns, and that the learned background noise profile provides evidence for degraded or uncaptured cell types.