SpaHNR: a spatial domain identification method via sparse attention-based hierarchical node representation and multi-view contrastive learning.

backgroundLeveraging deep learning on spatial transcriptomics data enables the identification of distinct spatial domains within tissues, thereby clarifies the spatial organization of cells and their gene expression. This detailed spatial understanding is crucial for dissecting intricate biological processes and the mechanisms of disease. However, existing computational approaches often analyze cell or spot (representing a small number of cells) adjacency relationships at a single level of detail (resolution), overlooking the natural hierarchical organization inherent in biological tissues.

resultsIn this study, we develop a novel method called SpaHNR for spatial domain identification that acknowledges and utilizes the multi-layered structure of tissues by leveraging Sparse Attention-based Hierarchical Node Representation and multi-view contrastive learning. SpaHNR treats each spot as a node and constructs two distinct spot views by integrating tissue images, gene expression profiles, spatial coordinates, and inferred cell communications. Subsequently, in each view, a single-layer Graph Convolutional Network (GCN) is applied to aggregate neighbor information and further enhances the spot features. The outputs of the GCNs are then fed into a sparse attention-based hierarchical node fusion module to generate coarse-grained node representations, which aim to capture multi-scale feature information of the spots. Finally, the fine-to-coarse node assignment matrix is decoded to reconstruct the augmented gene expression matrix. During model training, a combination of gene expression reconstruction loss and cross-view contrastive loss is used to optimize the model parameters. Spatial domains are ultimately delineated using the Leiden clustering algorithm based on the fine-to-coarse node assignment matrix features. Evaluation on the human dorsolateral prefrontal cortex dataset, human breast cancer dataset, and mouse embryo dataset demonstrates that SpaHNR outperforms existing state-of-the-art methods in most cases.

conclusionSpaHNR effectively integrates spatial location, gene expression profiles, tissue image information, and inferred cell communication to enhance spatial domain identification performance through hierarchical representation learning and contrastive learning strategies applied to spatial transcriptomics data. This demonstrates its significant potential for applications in the analysis of complex tissue architectures.