Focused single-cell analysis with principal feature analysis, mutual information, and machine learning reveals cell type signatures

Abstract Our single-cell analysis pipeline identifies a set of key genes whose expression describes the difference between cell types and consists of (i) R packages Seurat and DoubletFinder preprocess single cell data according to quality. Next, a custom-made python script randomizes obtained data, generating random sets for comparison. (ii) Principal feature analysis (PFA) identifies genes contributing to characteristic differences in cell types, represented by corresponding labels, removing redundant genes that are functions of others, leaving only genes that are arguments of such functions. (iii) Genes associated with predicting the labels of the cells are selected. For further reduction, selected genes are ranked by the mutual information that the expression of these genes shares with the labels representing the cell types. (iv) Importantly, selected genes are validated by a machine learning framework where the prediction accuracy on all genes vs. the selected genes is tested. (v) If desired, mutual information ranking balances the number of selected genes and information content for predicting the cell types. Results from different single-cell data include: comparisons of adipocytes and adipose stem and progenitor cells (ASPCs), adipocytes and mesothelium, and endothelial cells and mesothelium, as well as fibroblast and macrophage. In each, only about ten out of more than 30000 genes are identified carrying the relevant information classifying the single cells for distinguishable cell types. Our code is provided in a GitHub repository at … ( link will be added after acceptance).