Variability in scRNA-Seq analysis is mitigated by regularized or supervised approaches

Abstract

Transcriptional dynamics of evolutionary processes through time are highly complex and require single-cell resolution datasets. This is especially important in cancer during the evolution of resistance, where stochasticity can lead to selection for divergent transcriptional mechanisms. Statistical methods developed to address various questions in single-cell datasets are prone to variability and require careful adjustments of multiple parameter space. To assess the impact of this variation, we utilized commonly used single-cell RNA-Seq analysis tools in a combinatorial fashion to evaluate how repeatable the results are when different methods are combined. In the context of clustering and trajectory estimation, we benchmark the combinatorial space and highlight ares and methods that are sensitive to parameter changes. We have observed that utilizing temporal information in a supervised framework or regularization in latent modeling reduces variability leading to improved overlap when different parameters/methods are used. We hope that future studies can benefit from the results presented here as use of scRNA-Seq analysis tools as out of the box is becoming a standard approach in cancer research.