sigRGCN: A Robust Residual Graph Convolutional Network for scRNA-Seq Data Clustering.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Clustering is vital for single-cell RNA sequencing (scRNA-seq) analysis, enabling cell type discovery.
• Graph convolutional networks (GCNs) are effective for scRNA-seq clustering but suffer from noise sensitivity and over-smoothing.
• These limitations hinder the accurate capture of cell-specific information and higher-order relationships.

Purpose of the Study:

• To develop a robust graph convolutional network model for improved scRNA-seq data clustering.
• To address the challenges of noise sensitivity and over-smoothing inherent in existing GCN methods for scRNA-seq analysis.
• To enhance the accuracy and reliability of cell type identification through advanced clustering techniques.

Main Methods:

• Proposed sigRGCN, a residual graph convolutional network incorporating graph structure optimization.
• Constructed a disturbed cell graph with injected noise and employed a GCN to mitigate noise impact.
• Utilized an L-layers residual GCN to combat over-smoothing and capture higher-order cell relationships.
• Optimized the model using a self-supervised learning approach.

Main Results:

• The proposed sigRGCN model demonstrated significant robustness in clustering real-world scRNA-seq data.
• Effectively alleviated the over-smoothing problem, leading to improved cell representations.
• Achieved competitive performance across nine diverse scRNA-seq datasets.
• Successfully captured higher-order cellular relationships, enhancing clustering accuracy.

Conclusions:

• sigRGCN offers a robust and effective solution for scRNA-seq data clustering.
• The model successfully overcomes key limitations of traditional GCNs in this domain.
• Results indicate sigRGCN's potential for advancing cell type discovery and biological insights from scRNA-seq data.