CanSig: discovery of shared transcriptional states across cancer patients from single-cell RNA sequencing data

Abstract

Human tumors are highly heterogeneous in their cell composition; specifically, they exhibit heterogeneity in transcriptional states of malignant cells, as has been recently discovered through single-cell RNA sequencing (scRNA-seq). Distinct states of malignant cells have been linked to variability in tumorigenic properties and resistance to anti-cancer treatment. Despite the fact that scRNA-seq data contain all necessary information to uncover shared transcriptional states of malignant cells in tumors, jointly analyzing cells from multiple cancer patients comes with its set of challenges including batch correction and accounting for patient-specific genetic background driving differences between gene expression vectors. We propose CanSig, an easy-to-use approach designed to discover known and de novo shared signatures in cancer single cells. CanSig preprocesses, integrates and analyzes scRNA-seq data to provide new signatures of shared transcriptional states and links these states to known pathways. We show that CanSig successfully rediscovers ground truth pathways determining shared transcriptional states in two simulated and three experimental datasets; the latter spanning 135 patients and 72,000 cells. We then illustrate CanSig's investigative potential by discovering novel signatures in esophageal squamous cell carcinoma possibly linked to targeted patient treatment; we also point out a de novo signature in breast cancer predictive of patients' survival. In the cancer types studied, we juxtapose copy number variation with discovered shared transcriptional states and uncover a genetic component predisposing cancer cells to activation of specific transcriptional programs. In sum, CanSig, specifically developed to analyze shared transcriptional heterogeneity of malignant cells of different genetic backgrounds, can greatly facilitate the exploratory analysis of scRNA-seq cancer data and efficiently identify novel transcriptional signatures linked to known biological pathways.