The advent of single-cell RNA sequencing (scRNA-seq) technology has enabled the analysis of cellular heterogeneity at the single-cell level. In scRNA-seq data analysis, cell clustering is a crucial downstream task, as it facilitates the discovery of novel cell subtypes and the identification of known cell types, laying the groundwork for subsequent analyses. However, scRNA-seq data pose significant challenges for cell clustering due to their high dimensionality, sparsity, and technical artifacts (including batch effects and dropout events). To address these challenges, we propose GCLSC (Graph Contrastive Learning for Single-Cell Clustering), a novel graph contrastive learning model. GCLSC integrates Graph Transformer and Graph Attention Network (GAT) to co-model local cellular interactions and global topological dependencies. Four data augmentation strategies enhance data diversity and mitigate overfitting. Experiments on nine real-world scRNA-seq datasets demonstrate that the model achieves superior clustering accuracy. Furthermore, leveraging its excellent clustering performance, this architecture provides a reliable computational tool for cell population profiling of scRNA-seq data — its accurate clustering results can effectively support core tasks such as novel cell subtype identification and known cell type annotation. GCLSC exemplifies the potential of combining GAT, Transformer, and contrastive learning for robust single-cell analysis. The source code is available at https://github.com/JianjunTan-Beijing/GCLSC.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号