Abstract
The spatial transcriptomics technique provides an unprecedented perspective for analyzing the distribution patterns of cells within tissues and their functional tissue structures. To enhance the accuracy and robustness of spatial domain identification, we propose Joint Graph-Regularized Non-negative Matrix Factorization (JGR-NMF). An adaptive neighborhood graph construction strategy is introduced by applying an nth-power transformation to the spot adjacency probability matrix, thereby automatically optimizing the neighborhood size for individual spots. Furthermore, a JGR-NMF framework is developed, integrating this adaptively constructed kNN graph with the spatial adjacency matrix. Evaluations conducted on two breast cancer datasets, one Mouse Kidney dataset and one Mouse Embryo dataset, demonstrate that JGR-NMF significantly outperforms five state-of-the-art baseline methods in spatial domain identification. Systematic ablation studies further confirm the critical role of graph regularization in enhancing model performance.