scGGC: a two-stage strategy for single-cell clustering through cellular gene pathway construction

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) data analysis faces challenges due to high dimensionality, noise, and sparsity.
• Conventional clustering methods often overlook global cell-gene interactions, limiting accuracy.
• Advanced clustering algorithms are crucial for unlocking the potential of scRNA-seq data.

Purpose of the Study:

• To develop a novel single-cell clustering model, scGGC, that addresses limitations of existing methods.
• To improve the accuracy and biological relevance of clustering in scRNA-seq data.
• To enhance the identification of cell-type-specific marker genes.

Main Methods:

• scGGC integrates graph autoencoders and generative adversarial networks for single-cell clustering.
• An adjacency matrix captures both cell-cell and cell-gene relationships for graph structure construction.
• Nonlinear dimensionality reduction and initial clustering are performed via a graph autoencoder, followed by adversarial training for performance enhancement.

Main Results:

• scGGC demonstrated superior performance across nine scRNA-seq datasets compared to eight existing methods.
• Adjusted Rand Index improved by an average of 10.1% on datasets like MHC3K.
• Marker gene identification and cell type annotation showed biological relevance, with overlap rates exceeding 70%.

Conclusions:

• scGGC significantly improves the accuracy of single-cell data clustering.
• The model enhances the discovery of biologically relevant cell-type-specific marker genes.
• scGGC offers a robust approach for analyzing complex scRNA-seq data.