Statistical significance of cluster membership for determination of cell identities in single cell genomics

Abstract

Single cell RNA sequencing (scRNA-seq) allows us to dissect transcriptional heterogeneity arising from cellular types, spatio-temporal contexts, and environmental stimuli. Cell identities of samples derived from heterogeneous subpopulations are routinely determined by clustering of scRNA-seq data. Computational cell identities are then used in downstream analysis, feature selection, and visualization. However, how can we examine if cell identities are accurately inferred? To this end, we introduce non-parametric methods to evaluate cell identities by testing cluster memberships of single cell samples in an unsupervised manner. We propose posterior inclusion probabilities for cluster memberships to select and visualize samples relevant to subpopulations. Beyond simulation studies, we examined two scRNA-seq data - a mixture of Jurkat and 293T cells and a large family of peripheral blood mononuclear cells. We demonstrated probabilistic feature selection and improved t-SNE visualization. By learning uncertainty in clustering, the proposed methods enable rigorous testing of cell identities in scRNA-seq.