Spatial multi-omic technologies enable simultaneous multi-omic profiling within native tissue context, offering unprecedented opportunities to study biological processes and disease. As investigations of tissue spatial architecture gain in complexity, broadly applicable models are required to support vertical (within a section), horizontal (across sections), and mosaic (across distinct omics) integration. Here, we propose SpaMode, a broadly applicable framework designed to accommodate spatial multi-omic integration for all three modes and four omic types, including transcriptomics, proteomics, epigenomics, and metabolomics. SpaMode disentangles each omics modality into modality-invariant and modality-variant distributions to characterize underlying biomolecular commonalities and specificities, and then hierarchically aggregates these distributions to resolve spatial heterogeneity. Horizontal and mosaic integration are unified within the SpaMode framework through multi-slice joint regularization and translation of missing modalities. We benchmark SpaMode across vertical, horizontal, and mosaic integration, which demonstrate that SpaMode outperforms existing, targeted approaches in all integration settings. Furthermore, SpaMode provides novel insights into how invariant and variant multi-level biomolecular features contribute divergently to tissue spatial context, offering an interpretable alternative to black-box neural network. SpaMode provides a general and trustworthy solution for spatial multi-omic data analysis, paving the way for systematically decoding the complex mechanisms of cellular states and disease evolution in situ.
Zheng et al. (Mon,) studied this question.