Diversity in transcriptomics without cell types

Abstract

Downstream analysis in single-cell and spatial transcriptomics is highly dependent on a sequence of upstream modeling choices. The non-canonicity of these choices presents challenges for reproducibility. In particular, measures of cellular heterogeneity and diversity do not solely reflect biological variation, but are also sensitive to parameter settings. A diversity measure that is robust to modeling choices, such as clustering resolution, is therefore desirable to improve reproducibility and interpretability. Here, we introduce scDIV, a similarity-sensitive measure of cellular diversity inspired by mathematical ideas in ecological science, which is robust to graph-based clustering parameters and remains applicable even in the absence of cell-type clusters. We use scDIV to quantitatively track the progress of tissue differentiation in both single-cell and spatial mouse development datasets and to evaluate different engineered stem-cell-based embryo models. In contrast to traditional entropy-based methods, such as the Hill number, used to quantify biodiversity, scDIV remains robust to clustering.