SpaHDSRL: hierarchical dual-graph self-supervised representation learning for integrating spatially resolved...

Spatial multi-omics technologies facilitate simultaneous measurement of multiple molecular modalities within their native spatial context, offering opportunities to characterize tissue organization and cellular heterogeneity. However, effective integration remains challenging because such data concurrently encode spatial adjacency and molecular similarity, while also being limited by high dimensionality, sparsity, noise, and cross-modality heterogeneity. Here, we propose SpaHDSRL, a hierarchical dual-graph self-supervised representation learning framework for spatial multi-omics integration. SpaHDSRL jointly models a shared spatial graph and modality-specific feature graphs, which are integrated through an adaptive gated hierarchical fusion strategy to learn coherent and informative latent representation. To further enhance representation quality, SpaHDSRL combines a Deep Graph Infomax-based objective with spatial regularization, preserving both global informativeness and local spatial consistency. Experiments on simulated and real datasets demonstrate that SpaHDSRL consistently achieves superior performance over existing methods in both the accuracy and robustness of spatial domain identification. Downstream analyses further highlight its utility in marker discovery, functional enrichment, second-modality-associated analysis, and cell-cell communication inference, underscoring its value for dissecting tissue architecture, developmental programs, and multicellular interactions in complex biological systems. The source code of SpaHDSRL is available at https://github.com/Lisa62103/SpaHDSRL.