Label-Guided Graph Contrastive Learning for Single-Cell Fusion Clustering

Single-cell RNA sequencing (scRNA-seq) technology provides gene expression information at the individual cell level and reveals cellular heterogeneity within tissues. Cell clustering is an important task in scRNA-seq data analysis. Although many single-cell clustering methods have been proposed, existing methods often fail to fully consider both cell attribute information and the structural relationships between cells. Moreover, many graph clustering methods combine with contrastive learning, but most graph contrastive learning methods overlook the similarity between nodes. To address this, a label-guided graph contrastive learning-based single-cell fusion clustering method, scLGGCL, is proposed. First, a dual-reconstruction information fusion module is constructed to extract both the latent attribute information and the relationships between cells, thus obtaining a fusion of attribute and structural information. Next, a label-guided graph contrastive learning module is designed to capture semantic-level feature similarity between nodes and adjust the distance between positive and negative nodes based on predicted label information. Finally, a deep embedding clustering-based self-optimization module is introduced, which utilizes the fused attribute and structural information to optimize the clustering results and pull cells toward the cluster centers. The validity and accuracy of scLGGCL clustering were verified by comparing with other single-cell clustering methods on both single datasets and cross-datasets. The source code of scLGGCL is available at https://github.com/CDMBlab/scLGGCL.