scDCL: A multi-view single-cell RNA sequencing clustering method based on dual contrastive learning.

Single-cell RNA sequencing (scRNA-seq) has become a pivotal tool for resolving cellular heterogeneity and uncovering novel cell states. While deep learning-based scRNA clustering algorithms have significantly enhanced cellular resolution and computational efficiency, they still face persistent challenges inherent in scRNA-seq data: high sparsity, strong nonlinearity, and extreme dimensionality. Existing methods face a critical trade-off: autoencoder-based approaches capture cell-intrinsic features but neglect global inter-cell relationships, while graph neural network (GNN) or contrastive learning-based methods focus on structural information but lose cell-specific characteristics. To address this, we propose scDCL, a novel scRNA-seq clustering framework integrating ZINB-based masked autoencoders (MAE), GNNs, and dual contrastive learning to synergistically capture both cell-intrinsic and global structural features. It first uses ZINB-MAE to denoise data and learn initial representations, then constructs multiple graphs to model inter-cell relationships. Laplacian filtering integrates these graphs with raw data to generate multi-perspective features balancing local and global information. These multi-view features are input to GNNs to generate four embeddings, optimized via dual contrastive learning with intra-representation loss that enhances cluster compactness and separation and inter-representation loss that ensures structural consistency. The final clustering results are obtained by performing clustering on the optimized integrated representations. Extensive experiments on public datasets demonstrates that our integrated framework achieves superior clustering performance compared to state-of-the-art methods.