Contrastive graph regularized non-negative matrix factorization for domain identification of spatial transcriptomics

Spatial transcriptomics captures gene expression with spatial resolution, but its high dimensionality complicates spatial domain identification. While deep learning excels in feature extraction, its limited interpretability underscores the need for dimensionality reduction techniques that preserve spatial and biological relevance. In this study, we propose a novel contrastive graph-regularized non-negative matrix factorization (CGNMF) model for interpretable dimensionality reduction in spatial transcriptomics analysis. Our approach integrates graph regularization with a self-supervised contrastive learning framework to enhance both feature representation and spatial structure preservation. Specifically, we construct positive and negative sample pairs by jointly considering gene expression similarity and spatial proximity, enabling the model to learn discriminative representations that reflect both transcriptomic and spatial characteristics. The contrastive learning component is incorporated into the graph-regularized non-negative matrix factorization framework, effectively guiding the factorization process towards biologically and spatially coherent dimensions. This integration facilitates the automatic delineation of spatial domains and improves interpretability. We benchmark CGNMF against seven spatial domain identification methods using three publicly available datasets. Evaluations based on clustering metrics showed that CGNMF consistently outperformed existing methods. Notably, CGNMF successfully identified biologically relevant functional regions that are overlooked by current approaches, highlighting its robustness and utility in spatial domain identification tasks.