Local-Global Graph Fusion to Enhance scRNA-Seq Clustering

Single-cell RNA sequencing (scRNA-seq) is crucial for demystifying the cell heterogeneity and differentiation processes, enabling the identification of distinct cell subtypes within a population. However, most of the existing approaches are feeble to comprehensively investigate the interactive relationships between cells and exploit the topological structures of the scRNA-seq data, resulting in the accurate identification of cell types hard to ploughed. In this paper, we propose scLGF, a novel scRNA-seq deep clustering model with Local and Global Graph Fusion. Specifically, scLGF first generates a latent representation for each cell using the dual embedding learning module. Then, scLGF introduces a local and global graph fusion module to effectively capture underlying connections between cells to enhance the model's representative capabilities. Finally, scLGF proposes an optimized triplet graph self-supervised learning approach to learn the discriminative feature representations of cells. We use the fused consensus representation to generate reliable target distributions to supervise the dual embedding learning task. In this way, the three modules can mutually enhance each other end-to-end. Experimental results demonstrate the superiority of scLGF over six alternative methods on ten widely used single-cell datasets. Moreover, scLGF exhibits scalability on large-scale datasets, making it a practical tool for scRNA-seq data analysis. The source codes are available online at https://github.com/lijing2000/scLGF.