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

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity but faces challenges like sparsity and high dimensionality.
• Current deep learning clustering methods often struggle to balance cell-intrinsic features with global inter-cell relationships.

Purpose of the Study:

• To develop an integrated framework, scDCL, that synergistically captures both cell-intrinsic and global structural features for improved scRNA-seq clustering.
• To address the limitations of existing methods that either focus on cell-specific characteristics or global structural information.

Main Methods:

• Proposed scDCL framework integrating Zero-Inflated Negative Binomial-based Masked Autoencoders (ZINB-MAE), Graph Neural Networks (GNNs), and dual contrastive learning.
• Utilized ZINB-MAE for denoising and initial representation learning, followed by graph construction and Laplacian filtering for multi-perspective feature generation.
• Employed GNNs and dual contrastive learning to optimize integrated representations, enhancing cluster compactness, separation, and structural consistency.

Main Results:

• The scDCL framework demonstrated superior clustering performance on public scRNA-seq datasets compared to existing state-of-the-art methods.
• Achieved a balance between capturing cell-intrinsic features and global structural information, overcoming limitations of previous approaches.

Conclusions:

• scDCL offers a powerful and integrated approach for scRNA-seq data analysis, significantly improving cell clustering accuracy.
• The proposed framework provides a robust solution for resolving cellular heterogeneity and uncovering novel cell states in complex biological systems.