scDFVA: Single-Cell Deep Clustering by Fusing Variational Graph Attention Autoencoder and ZINB-Based Autoencoder

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables high-resolution cellular analysis.
• Accurate cell type identification via single-cell clustering is essential for understanding biological functions.
• Existing deep learning models struggle with the high dimensionality and sparsity of scRNA-seq data.

Purpose of the Study:

• To develop a novel deep clustering model, scDFVA, for improved scRNA-seq data analysis.
• To effectively capture intrinsic attributes and structural relationships within complex scRNA-seq datasets.
• To enhance the accuracy of cell type identification and functional analysis.

Main Methods:

• Proposed scDFVA model integrating a variational graph attention autoencoder (AE) and a zero-inflated negative binomial (ZINB) based AE.
• Incorporated ZINB model to better simulate sparse and zero-inflated scRNA-seq data.
• Employed self-supervised clustering on latent fusion representations for mutual supervision between representation learning and clustering.

Main Results:

• scDFVA effectively fuses gene expression and cell structure information within a joint framework.
• The model demonstrated superior performance compared to several competing methods in single-cell clustering tasks.
• Experimental results validate the efficacy of scDFVA in analyzing complex scRNA-seq data.

Conclusions:

• scDFVA offers a robust and effective approach for single-cell deep clustering.
• The proposed method enhances the ability to reveal cell types and their functions from scRNA-seq data.
• scDFVA represents a significant advancement in computational approaches for single-cell genomics.