Abstract
BACKGROUND: Graph Convolutional Networks (GCNs) are widely applied for spatial domain identification in spatial transcriptomics (ST), where node representations are learned by aggregating information from neighboring spots. However, most ST workflows construct spatial graphs by assigning equal weights to neighbors and self-loops, and then applying degree-based normalization. This procedure often yields near-uniform adjacency matrices, suppressing natural distance heterogeneity, diminishing spatial resolution, aggravating GCN over-smoothing, and obscuring fine-grained tissue boundaries. METHODS: We introduce DWGCN, a Distance-Weighted Graph Convolutional Network that replaces uniform neighbor assignment with inverse-distance weighting (IDW) and spot-wise normalization. DWGCN enhances locality-sensitive aggregation by assigning larger weights to proximal neighbors, while preserving self-loop dominance to maintain intrinsic spot information and reduce hub-driven dilution. RESULTS: Across four real and four simulated ST datasets, integrating DWGCN with representative GCN-based frameworks (SEDR, GraphST, SpaNCMG, SpaGIC) generally improved clustering accuracy, particularly in tissues with complex spatial architectures. CONCLUSION: These results demonstrate that DWGCN offers a broadly applicable approach for restoring distance-aware structure in spatial graphs, thereby improving the delineation of spatial domain identification.