Demystifying "drop-outs" in single cell UMI data

Abstract

Analysis of scRNA-seq data has been challenging particularly because of excessive zeros observed in UMI counts. Prevalent opinions are that many of the detected zeros are "drop-outs" that occur during experiments and that those zeros should be accounted for through procedures such as normalization, variance stabilization, and imputation. Here, we extensively analyze publicly available UMI datasets and challenge the existing scRNA-seq workflows. Our results strongly suggest that resolving cell-type heterogeneity should be the foremost step of the scRNA-seq analysis pipeline because once cell-type heterogeneity is resolved, "drop-outs" disappear. Additionally, we show that the simplest parametric count model, Poisson, is sufficient to fully leverage the biological information contained in the UMI data, thus offering a more optimistic view of the data analysis. However, if the cell-type heterogeneity is not appropriately taken into account, pre-processing such as normalization or imputation becomes inappropriate and can introduce unwanted noise. Inspired by these analyses, we propose a zero inflation test that can select gene features contributing to cell-type heterogeneity. We integrate feature selection and clustering into iterative pre-processing in our novel, efficient, and straightforward framework for UMI analysis, HIPPO (Heterogeneity Inspired Pre-Processing tOol). HIPPO leads to downstream analysis with much better interpretability than alternatives in our comparative studies.