spaDWGCN: a spatial distance weighted graph convolution network for accurate spatial tissue structure identification

The continuous development of spatial transcriptomics enables us to obtain gene expression matrices alongside spatial coordinate labels for cells or spots, providing tools for understanding spatial tissue microenvironments and spatially resolved cell-cell interactions. Although numerous methods have been developed to identify spatial tissue architecture, most of them adopt binary adjacency matrices that treat all neighboring cells equally, and few comprehensively incorporate distance information into both graph construction and feature representation. Here we introduce spaDWGCN, a spatial distance weighted graph convolutional network algorithm that takes cellular gene expression and microenvironment composition as inputs. This approach introduces fine-grained distance weighting at two levels: replacing the conventional 0-1 binary adjacency matrix with a distance-weighted adjacency matrix at the graph level, and substituting the neighbor count matrix with a distance-weighted microenvironment composition matrix at the feature level. Furthermore, spaDWGCN jointly models gene expression and microenvironment composition through separate graph convolutional network branches and trainable fusion, integrating transcriptional patterns with their local neighborhood context within a unified framework. Experiments on spatial transcriptomics multi-platform datasets, including MERFISH, STARmap, Stereo-seq, Slide-seqV2, and 10× Visium, demonstrate that this distance-weighted strategy achieves competitive performance in spatial domain identification and other downstream analyses.