Abstract
BACKGROUND: Single-cell RNA sequencing (scRNA-seq) allows the exploration of biological heterogeneity among different cell types within tissues at a single-cell resolution. Cell clustering serves as a foundation for scRNA-seq data analysis and provides new insights into the heterogeneity of cells within complex tissues. However, the inherent features of scRNA-seq data, such as heterogeneity, sparsity, and high dimensionality, pose significant technical challenges for effective cell clustering. RESULTS: Here, we present a novel deep clustering method, scCAGN, based on an adversarial autoencoder (AAE) and a cross-attention graph convolutional network (GCN), to address the above challenges in scRNA-seq data analysis. Specifically, to enhance data reconstruction, scCAGN utilizes adversarial autoencoders to augment encoder capabilities. Graph feature representations obtained via a GCN were integrated using a dynamic information fusion mechanism, yielding enhanced feature representations. In addition, scCAGN combines three different loss functions to optimize clustering performance through a joint clustering approach. By leveraging a unique information fusion and joint mechanism, scCAGN extracts deep cell features without labeled information, thus improving cell classification efficiency. Our findings show that scCAGN surpasses the existing methods in clustering performance across eight typical scRNA-seq datasets, achieving a maximum Normalized Mutual Information (NMI) improvement of 11.94%, notably reaching an NMI of 0.9732 in the QS_diaphragm dataset. It showed an average NMI improvement of 13% across the eight benchmark datasets, surpassing the lowest-performing method. Further ablation and hyperparameter analyses validated the robustness of the proposed method. The code is available at: http://github.com/gladex/scCAGN . CONCLUSION: scCAGN integrates AAE and cross-attention GCN with dynamic fusion, achieving state-of-the-art scRNA-seq clustering (0.9732 NMI, 13% average gain) across eight datasets. Validated via ablation and hyperparameter tests, it advances label-free cell discovery and enables further multimodal integration to dissect cellular heterogeneity.