Self-supervised Graph Contrastive Learning for scRNA-seq Clustering

Single-cell RNA sequencing (scRNA-seq) technology has enabled us to characterize cellular heterogeneity, and analyzing scRNA-seq data can help enhance our understanding of complex diseases. Clustering of scRNA-seq data is a critical step in the analysis, as it assigns cells to subpopulations, facilitating a deeper understanding of cellular diversity. Although numerous scRNA-seq clustering algorithms have been proposed, they often fail to consider inherent cell type information and the intrinsic relationships between cells. As a result, the learned representations may be suboptimal for accurate cell type identification, limiting overall performance. To address these challenges, we propose a self-supervised graph learning (SSGL) method for scRNA-seq clustering. SSGL aims to enhance feature representation learning and achieve more stable clustering results by incorporating both inherent cell type information and cell relationships. Specifically, our method applies dual random masking to gene expression profiles, generating two augmented datasets. We assume that samples within the same cluster, along with their augmentations, should exhibit similarity. To enforce this assumption, we construct a graph that captures inherent cell relationships across the augmented datasets. Additionally, high-confidence cell type identifications are leveraged to guide representation learning and ensure consistency between the clustering results of the two augmented datasets. This approach strengthens feature representation robustness and improves clustering stability. Extensive experiments on multiple public datasets demonstrate that our method outperforms baseline algorithms in clustering accuracy and provides valuable biological insights into scRNA-seq data.