SpatialGlue is a novel deep learning method for integrating spatial multi-omics data in a spatially informed manner. It utilizes a cycle graph neural network with a dual-attention mechanism to learn the significance of each modality at cross-omics and intra-omics integration. The method can accurately aggregate cell types or cell states at a higher resolution on different tissue types and technology platforms. Besides, it can provide interpretable insights into cross-modality spatial correlations. SpatialGlue is computationally efficient and it only requires about 5 mins for spatial multi-omics data at single-cell resolution (e.g., Spatial-ATAC-RNA-seq data, ~10,000 spots).