An information theoretic approach to detecting spatially varying genes

Abstract

Identifying genes with spatially coherent expression patterns is a key task in spatial transcriptomics. We adopt an information theoretic perspective on this problem by equating the degree of spatial coherence with the mutual information between nearby expression measurements. To avoid the notoriously difficult problem of computing mutual information, we use modern methods of approximation, in a method we call maximization of spatial information (Maxspin). As well as being highly scalable, we demonstrate improved accuracy across several spatial transcriptomics platforms and a variety of simulations when compared to both existing specialized methods and traditional spatial statistics methods. We use the method to analyze a renal cell carcinoma sample profiled using CosMx Spatial Molecular Imaging, revealing previously undescribed gene expression patterns.