Abstract
When cells are isolated for single-cell RNA sequencing (scRNA-seq), their positional information is inevitably lost. Here, an algorithm termed MERLIN is described that can reconstruct such information in organs with compartmentalized anatomy, such as the kidney. Several independent immune cell scRNA-seq datasets from three renal compartments were generated to train different machine learning algorithms. A modified multi-layer Perceptron approach most accurately predicted positions of resident macrophages, best, achieving over 75% accuracy in both murine and human kidney datasets. More motile immune cells, like lymphocytes, were not predictable. Positional transcriptomic fingerprints were enriched in pathways of microenvironmental responses and cellular adaptation, and showed a sex bias. MERLIN also predicted positions of resident and recently recruited macrophages in a crescentic glomerulonephritis mouse model. Analysis of published scRNA-seq datasets from endotoxin- and ischemia/reperfusion-induced models of acute kidney injury revealed proinflammatory responses predominantly in outer medullary macrophages, consistent with the known pathology. Moreover, the response of cortical macrophages to commonly used therapies for diabetic nephropathy aligned with the known clinical drug efficacy. Finally, MERLIN was successfully trained to predict the spatial distribution of brain microglia. Together, MERLIN enables spatial interpretation of scRNA-seq datasets in organs with defined anatomical regions and enhances mechanistic insights into disease processes.