Statistically principled feature selection for single cell transcriptomics

The high dimensionality of data in single cell transcriptomics (scRNAseq) requires investigators to choose subsets of genes (feature selection) for downstream analysis (e.g., unsupervised cell clustering). The evaluation of different approaches to feature selection is hampered by the fact that, as we show here, the performance of feature selection methods varies greatly with the task being performed. For routine cell type identification, even randomly chosen features can perform well, but for cell type differences that are subtle, both number of features and selection strategy can matter strongly. Here we present a simple feature selection method grounded in an analytical model that, without resorting to arbitrary thresholds or user-defined parameters, allows for interpretable delineation of both how many and which features to choose, facilitating identification of biologically meaningful rare cell types. We compare this method to default methods in scanpy and Seurat, as well as SCTransform, showing how greater accuracy can often be achieved with surprisingly few, well-chosen features.