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

Given the growing number of clinically integrated cancer single-cell transcriptomic studies, robust differential enrichment methods for gene signatures to dissect tumor cellular states for discovery and translation are critical. Current analysis strategies neither adequately represent the hierarchical structure of clinical single-cell transcriptomic datasets nor account for the variability in the number of recovered cells per sample, leading to results potentially confounded by sample-driven biology instead of accurately representing true differential enrichment of group-level biology (e.g., treatment responders vs. non-responders) and a high number of false positives. This problem is even more severe for single-cell analyses of the tumor compartment due to high intra-patient similarity (as opposed to inter-patient similarity), leading to stricter hierarchical structured data for this compartment. Furthermore, to identify signatures which are truly representative of the entire group, there is a need to quantify the robustness of otherwise statistically significant signatures to sample exclusion. Here, we present a new nonparametric statistical method, BEANIE, to account for these issues, and demonstrate its utility in two cancer cohorts stratified by clinical groups to reduce biological hypotheses and guide translational investigations. Using BEANIE, we show how the consideration of sample-specific versus group biology helps decrease the false positive rate by more than 10 times and identify robust signatures which can also be corroborated across different cell type compartments.