Learning tissue representation by identification of persistent local patterns in spatial omics data

Abstract

Abstract Spatial omics data provide rich molecular and structural information on tissues. Their analysis provides insights into local heterogeneity of tissues and holds promise to improve patient stratification by associating clinical observations with refined tissue representations. We introduce Kasumi, a method for identifying spatially localized neighborhood patterns of intra- and intercellular relationships that are persistent across samples and conditions. The tissue representation based on these patterns can facilitate translational tasks, as we show for stratification of cancer patients for disease progression and response to treatment using data from different experimental platforms. On these tasks, Kasumi outperforms related approaches and offers explanations of spatial coordination and relationships at the cell-type or marker level. We show that persistent patterns comprise regions of different sizes, and that non-abundant, localized relationships in the tissue are strongly associated with unfavorable outcomes.