CellVGAE: an unsupervised scRNA-seq analysis workflow with graph attention networks.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-resolution data, necessitating advanced analytical methods.
• Deep learning approaches are increasingly vital for analyzing large-scale scRNA-seq datasets.

Purpose of the Study:

• To introduce CellVGAE, a novel variational graph autoencoder utilizing graph attention for unsupervised exploration of scRNA-seq data.
• To enhance dimensionality reduction, clustering, and interpretability of scRNA-seq data.

Main Methods:

• Development of a variational graph autoencoder (CellVGAE) with graph attention layers.
• Application of CellVGAE to analyze scRNA-seq data, focusing on connectivity between cells.
• Evaluation of model interpretability through graph attention coefficients.

Main Results:

• CellVGAE demonstrates superior interpretability compared to existing variational architectures by analyzing attention coefficients.
• The model effectively captures biological information like pseudotime and NF-κB activation dynamics.
• CellVGAE outperforms leading methods in dimensionality reduction and clustering on nine challenging datasets.
• Significant reduction in training times (up to 20x) observed on a large dataset.

Conclusions:

• CellVGAE provides an effective tool for exploratory analysis of scRNA-seq data, extracting meaningful features for visualization and interpretation.
• The model offers enhanced interpretability and performance, outperforming current deep learning alternatives.
• CellVGAE presents a computationally efficient solution for large-scale scRNA-seq data analysis.