Differential expression analysis in single-cell and spatial RNA-seq without model assumptions

Gene up(down)regulation findings in single-cell and spatial RNA sequencing can be inconsistent despite remarkable progress in technology. Inconsistent findings in high-quality samples raise concerns about assumptions behind widely accepted data analysis approaches. We, therefore, propose a weighted averaging approach for data analysis without assuming anything besides randomness of technical noise. This approach is closely related to prior work on statistics of cluster-randomized experiments. We show that weighing transcript counts based on measured noise variances and utilizing weighted rather than standard unweighted tests reduce both false-positive and false-negative findings. Our approach eliminates the need for parametrizing data distributions and/or rescaling transcript counts, which may cause artifacts by distorting and biasing the data. The resulting analysis is less complex and produces more consistent differential gene expression estimates.