Self-supervised graph contrastive learning for scRNA-seq clustering.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity but faces challenges in accurate clustering due to high dimensionality and noise.
• Existing methods often fail to fully utilize cell-cell relationships and cell-type signals, leading to suboptimal performance and unstable results.
• Developing robust and interpretable clustering methods is crucial for advancing scRNA-seq data analysis.

Purpose of the Study:

• To introduce a novel self-supervised graph contrastive framework for scRNA-seq clustering.
• To enhance clustering robustness and biological interpretability by integrating augmented views and graph refinement.
• To address limitations in existing scRNA-seq clustering techniques.

Main Methods:

• Proposed Self-Supervised Contrastive Graph Learning (SSGL) framework for scRNA-seq clustering.
• Utilized dual random gene masking for data augmentation and a momentum-encoder for representation learning.
• Constructed a refined cell-cell graph by combining k-nearest-neighbor similarity with pseudo-label consistency for graph-aware contrastive learning.

Main Results:

• SSGL demonstrated superior clustering performance across eight scRNA-seq benchmarks, achieving average NMI of 0.876 and ARI of 0.926.
• Outperformed state-of-the-art baselines, including AttentionAE-SC, with significant improvements in NMI (4.4%) and ARI (6.7%).
• Ablation studies confirmed the benefit of the self-supervised refined graph; visualization and marker-gene analyses validated biologically coherent cell group recovery, including rare populations.

Conclusions:

• SSGL significantly improves scRNA-seq clustering accuracy and stability.
• The framework effectively leverages augmented views, cell-cell relationships, and pseudo-label-guided graph refinement.
• SSGL provides robust representations for reliable cell-type discovery and downstream biological interpretation.