Deep learning powered single-cell clustering framework with enhanced accuracy and stability.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for understanding cellular diversity.
• Unsupervised clustering identifies cell types, but graph-based methods often ignore node distribution and suffer from oversmoothing.
• Existing methods struggle with accurate and scalable cell population representation.

Purpose of the Study:

• To introduce scG-cluster, a novel deep structural clustering method for scRNA-seq data.
• To address limitations of existing graph-based clustering, including node distribution neglect and oversmoothing.
• To improve the accuracy and scalability of cell type identification in scRNA-seq analysis.

Main Methods:

• Developed scG-cluster, featuring a dual-topology adjacency graph to incorporate node distribution.
• Employed a dual-topology adaptive graph convolutional network (TAGCN) with attention and residual connections.
• Implemented iterative refinement of clustering centers for enhanced stability.

Main Results:

• scG-cluster demonstrated superior clustering accuracy and scalability across six diverse scRNA-seq datasets.
• Ablation studies confirmed the effectiveness of the attention mechanism and residual connections.
• The method consistently outperformed state-of-the-art clustering approaches.

Conclusions:

• scG-cluster offers a robust and effective solution for unsupervised clustering in scRNA-seq data.
• The dual-topology graph and TAGCN architecture significantly improve cell population representation and differentiation.
• The proposed method advances the field of single-cell data analysis, providing more accurate and scalable cell type identification.