Dissecting tumor cell programs through group biology estimation in clinical single-cell transcriptomics.

Abstract

With the growth of clinical cancer single-cell RNA sequencing studies, robust differential expression methods for case/control analyses (e.g., treatment responders vs. non-responders) using gene signatures are pivotal to nominate hypotheses for further investigation. However, many commonly used methods produce a large number of false positives, do not adequately represent the patient-specific hierarchical structure of clinical single-cell RNA sequencing data, or account for sample-driven confounders. Here, we present a nonparametric statistical method, BEANIE, for differential expression of gene signatures between clinically relevant groups that addresses these issues. We demonstrate its use in simulated and real-world clinical datasets in breast cancer, lung cancer and melanoma. BEANIE outperforms existing methods in specificity while maintaining sensitivity, as demonstrated in simulations. Overall, BEANIE provides a methodological strategy to inform biological insights into unique and shared differentially expressed gene signatures across different tumor states, with utility in single-study, meta-analysis, and cross-validation across cell types.