We create data-driven maps of transcriptomic anatomy with a probabilistic framework for unsupervised pattern discovery in spatial gene expression data. With convolved negative binomial regression we discover patterns which correspond to cell types, microenvironments, or tissue components, and that consist of gene expression profiles and spatial activity maps. Expression profiles quantify how strongly each gene is expressed in a given pattern, and spatial activity maps reflect where in space each pattern is active. Arbitrary covariates and prior hierarchies are supported to leverage complex experimental designs. We demonstrate the method with Spatial Transcriptomics data of mouse brain and olfactory bulb. The discovered transcriptomic patterns correspond to neuroanatomically distinct cell layers. Moreover, batch effects are successfully addressed, leading to consistent pattern inference for multi-sample analyses. On this basis, we identify known and uncharacterized genes that are spatially differentially expressed in the hippocampal field between Ammon's horn and the dentate gyrus.

Charting Tissue Expression Anatomy by Spatial Transcriptome Decomposition