Fig. 1

Schematic overview of geneBasis. Below we describe the steps of the algorithm. (1) “True” k-NN graph is constructed, which represents the “ground truth” relationship between cells. Colors correspond to cell types for visualization but are not used by the algorithm. For each gene (Narf and Hba-a1 are used as examples), we calculate the Minkowski distance between two vectors: the first vector corresponds to log-normalized expression for each cell (referred to as measured); the second vector corresponds to the average log-normalized expressions across the first k neighbors for each cell (referred to as predicted). (2) For the current selection the “Selection” graph is constructed to represent the relationship between cells using currently selected genes. As in (1), for each gene we calculate the Minkowski distance between measured and predicted counts. (3) The Minkowski distances are compared for each gene, and the gene with the biggest difference between Selected and True graphs (in the schematic - Hba-a1) is added to the current selection. (4) Steps (2)–(3) are repeated until the number of selected genes reaches the desired value