Multi-scale microenvironment covariance for unsupervised spatial domain identification

Abstract

Motivation: Many spatial domain identification methods describe the neighbourhood of a cell by mean gene expression. However, the mean cannot separate neighbourhoods that differ only in gene-gene covariance and tissue is organised at several spatial scales. Results: We propose CoSMIM, which represents the niche of each cell by a weighted fusion of gene-gene covariance matrices of its spatial neighbourhood at multiple scales. Principal component analysis and Leiden clustering of the matrix square root of the fused matrix identify spatial domains without model training. On six benchmark datasets from five platforms, CoSMIM attains the best overall rank over eight evaluation metrics among 16 methods. It also ranks first on a Xenium mouse brain dataset, and on all seven datasets its median NMI exceeds that of single-scale COVET. The domain covariance matrices also show gene pairs whose correlation differs between domains.