CLGraph: A Contrastive Learning-Based Multimodal Fusion Approach for Spatial Domain Identification

Spatial transcriptomics technologies provide multi-modal data for exploring tissue architecture domains, cell types, intercellular communication, and their biological consequences, including high-throughput gene expression, spatial coordinates, and histological background. In spatial domain identification, spatial transcriptomics data reveal the positions and functions of cells within tissue structures. Existing spatial domain detection methods do not fully leverage the unique spatial properties of spatial transcriptomics data and the morphological features in corresponding tissue images. To integrate the complex relationships among multimodal data, this paper proposes an optimized representation model based on graph contrastive learning. It utilizes complementary information among different modalities to jointly optimize low-dimensional embeddings of gene expression profiles. The paper validates the performance of the proposed model in spatial domain identification across multiple datasets, demonstrating consistency with biological annotations. Additionally, spatially variable genes identified based on clustering results align with certain marker genes.