A hybrid adversarial autoencoder-graph network model with dynamic fusion for robust scRNA-seq clustering

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables high-resolution analysis of cellular heterogeneity.
• Effective cell clustering is crucial for scRNA-seq data analysis but challenged by data complexity.
• Existing methods struggle with the heterogeneity, sparsity, and high dimensionality of scRNA-seq data.

Purpose of the Study:

• To develop a novel deep clustering method for enhanced scRNA-seq data analysis.
• To address the technical challenges posed by scRNA-seq data characteristics.
• To improve the efficiency and accuracy of cell classification and discovery.

Main Methods:

• scCAGN employs an adversarial autoencoder (AAE) for enhanced data reconstruction.
• A cross-attention graph convolutional network (GCN) extracts robust graph feature representations.
• Dynamic information fusion and a joint clustering approach with three loss functions optimize performance.

Main Results:

• scCAGN achieved state-of-the-art clustering performance across eight scRNA-seq datasets.
• Demonstrated a maximum Normalized Mutual Information (NMI) improvement of 11.94% and an average gain of 13%.
• Achieved a high NMI of 0.9732 on the QS_diaphragm dataset, validating robustness through ablation and hyperparameter analyses.

Conclusions:

• scCAGN offers a robust and effective solution for scRNA-seq data clustering.
• The method advances label-free cell discovery and facilitates the dissection of cellular heterogeneity.
• It provides a foundation for future multimodal data integration in biological research.