Artificial intelligence identifies spatial fingerprints of kidney-resident macrophages and brain microglia

Abstract The isolation of immune cells from tissues for single cell RNA sequencing (scRNA-seq) loses spatial information. Here we employed machine learning methods to identify transcriptomic fingerprints containing such information. We used murine kidney and brain as organs with macroscopically distinguishable regions and generated separate scRNA-seq datasets of immune cells from distinct areas. Several machine learning algorithms were utilized to identify highly variable genes harboring spatial information. Multilayer perceptron (MLP) performed best at predicting the position of kidney-resident macrophages with an accuracy of >75%. No algorithm allowed predicting the position of motile immune cells like monocyte-derived macrophages or lymphocytes. Also kidney dendritic cell positions were not predictable, presumably because these were found to mostly belong to a recently described subset with monocyte lineage (DC3). The macrophage spatial fingerprints were enriched in pathways involved in microenvironmental responses and cellular adaptation, and they showed a gender bias. In an experimental crescentic glomerulonephritis, macrophage positioning was predicted with an accuracy of 69-71%. Also in a human dataset, our algorithm operated with comparable efficiency. Applying our strategy to brain datasets predicted microglia positions with 74% accuracy. Our approach to predict the location of tissue-resident macrophages may be applicable to organs other than kidney and brain.