SpatialDG: a novel spatial domain identification method for spatially resolved transcriptomics data based on dual-graph neural network

Spatially resolved transcriptomic (ST) technologies offer transformative opportunities to chart gene expression landscapes within intact tissue architecture. Uncovering spatially discrete domains of biological function is essential for deciphering tissue heterogeneity, developmental processes, and disease mechanisms. Yet, the inherent noise, high dimensionality, and spatial sparsity of ST data present substantial challenges to the unsupervised delineation of these domains. We present SpatialDG, a dual-graph self-supervised contrastive learning framework for ST. SpatialDG combines graph neural networks with self-supervised contrastive learning to learn informative and discriminative spot representations by maximizing the agreement between local (node) and global (graph) embeddings and leveraging spatial adjacency to enhance representation learning. Specifically, SpatialDG constructs both a gene expression similarity graph and a spatial adjacency graph, integrating them via a dual-view contrastive architecture that aligns molecular and spatial information, while a zero-inflated negative binomial reconstruction loss accounts for the count-based and sparse nature of gene expression data. SpatialDG achieves significant gains over state-of-the-art algorithms in both healthy and cancer datasets, demonstrating robust generalization across diverse ST landscapes. In conclusion, SpatialDG efficiently unravels biologically meaningful domains from spatial and genetic signals, providing a powerful and generalizable tool to mine tissue architecture in ST datasets.