Explainable machine learning approaches are well suited to flexibly and efficiently capture different aspects of organization in spatial omics data and offer a more detailed view of the underlying tissue biology. We developed MISTy, a scalable multi-view machine learning framework, to enable versatility of analysis by combining different types of complex spatially resolved data with prior knowledge. Our explainable models are used for data exploration and generation of hypothesis based on robust structural and functional relationship patterns in the data captured within different spatial contexts, ranging from the subcellular to the broader tissue context. With our newest development, Kasumi, we offer a novel perspective for capturing the heterogeneity of tissues by identification of spatially localized, persistent niches of relationships. Kasumi outperforms related approaches and offers new insights and explanations on the spatial coordination and multivariate relationships underlying the differences in observed progression and response in cancer.