Identifying spatial domains from spatial multi-omics data using consistent and specific deep subspace learning

The rapid advancement of spatial omics technologies has revolutionized the ability to simultaneously capture multi-omics data along with spatial information, offering unprecedented insights into tissue architecture and cellular heterogeneity. Spatial domain identification is a fundamental step in spatial omics analysis, as it facilitates the delineation of functional tissue regions. However, most existing methods are tailored for single omic data and face substantial limitations when extended to spatial multi-omics contexts. Integrating consistent and complementary signals across multiple omics within spatially structured data remains a key challenge. In this study, we propose SpaMICS, a deep subspace learning framework designed for spatial domain identification from spatial multi-omics data. SpaMICS captures spatial dependencies and latent inter-spot relationships to learn high-level representations for each omic. To enhance information integration, we introduce a subspace learning module that explicitly disentangles consistent and complementary information across omics. Furthermore, we incorporate dual constraints to enhance information extraction: a low-rank constraint to emphasize consistent information across omics and a discriminative constraint that facilitates the extraction of complementary information. Extensive experiments on five real-world spatial multi-omics datasets, including spatial transcriptomics–proteomics and spatial transcriptomics–epigenomics data, demonstrate that SpaMICS consistently outperforms existing approaches, effectively integrating multi-omics data for accurate spatial domain identification.