Single-cell spatial multi-omics sequencing technology can simultaneously obtain multiple omics data, such as transcriptome and epigenome, along with their spatial location information, providing a new perspective for deciphering cell-cell interactions and tissue microenvironments. Existing modeling methods rely on a single spatial scale and non-adaptive cross-modality fusion strategies, making it difficult to capture complex spatial heterogeneity. To address the above issues, this paper proposes a graph neural network model named Multi-Scale Adaptive Graph Convolutional Network, MSAGCN. The MSAGCN model constructs feature graphs and multi-scale spatial graphs by leveraging cell expression similarity and spatial location information, then extracts omics-specific embeddings through a dual-branch encoder, and finally obtains cross-scale and cross-modal consistent representations with the help of a spatial context attention mechanism. Experimental results on multiple spatial multi-omics datasets show that the MSAGCN model outperforms methods such as SpatialGlue and COSMOS in overall performance and also provides a new tool for spatial domain partitioning.
Li et al. (Thu,) studied this question.