SpatialFusion: A Unified Model for Integrating Spatial Transcriptomics to Unveil Cell-type Distribution, Interaction, and Functional Heterogeneity in Tissue Microenvironments.

Recent advances in spatial transcriptomics (ST) have significantly enhanced our understanding of tissue structure and intercellular interactions. However, existing methods for spatial domain identification and cell type deconvolution still face challenges related to accuracy, robustness, and computational efficiency. To address these issues, we introduce SpatialFusion, an innovative deep learning model designed to improve both spatial domain identification and cell type deconvolution by integrating gene expression and spatial coordinates. The core innovation of SpatialFusion lies in its use of graph neural networks (GNN) and attention mechanisms to capture complex spatial relationships through multi-dimensional embeddings of spatial data. By employing a dual-encoding strategy (co-learning of spatial graphs and feature maps) and self-supervised contrastive learning, the model significantly enhances accuracy and robustness across datasets. Experimental results demonstrate that SpatialFusion outperforms existing methods in accuracy and resolution when applied to the human DLPFC dataset, particularly in capturing complex, layer-specific expression patterns. The model also shows strong robustness in cell type deconvolution, accurately mapping spatial cell type distributions despite noise and low cell density. In breast cancer tumor microenvironment analysis, SpatialFusion revealed spatial heterogeneity and identified potential therapeutic targets, COX6C and CCND1, providing valuable insights for precision medicine.