CellVGAE

CellVGAE applies variational graph autoencoders and graph attention networks to perform unsupervised dimensionality reduction and clustering of single-cell RNA sequencing (scRNA-seq) data.

Key Features:

• Graph-Based Architecture: Constructs cell connectivity graphs such as k-nearest neighbor (KNN) graphs using gene expression values as node features to capture relationships between cells.
• Variational Graph Autoencoder: Encodes high-dimensional scRNA-seq data into a lower-dimensional latent space while preserving graph structural information.
• Graph Attention Networks (GAT): Integrates graph attention layers that compute attention coefficients to weight influential connections within the cell graph.
• Dimensionality Reduction and Clustering: Produces compact latent representations that facilitate visualization and clustering to identify distinct cell populations.
• Interpretability: Uses analysis of graph attention coefficients to highlight important nodes and edges contributing to learned representations.

Scientific Applications:

• Exploratory Analysis: Extracts meaningful features from scRNA-seq datasets to support exploratory investigation of cellular heterogeneity, including challenging datasets.
• Visualization and Interpretation: Enables visualization-driven interpretation of cellular relationships based on latent embeddings and attention-weighted graph structure.
• Comparative Performance: Demonstrated consistent outperformance of neural and non-neural methods for dimensionality reduction and clustering across nine well-annotated scRNA-seq datasets.

Methodology:

Uses unsupervised graph neural network methods—variational graph autoencoders combined with graph attention layers—applied to k-nearest neighbor (KNN) graphs with gene expression as node features to encode scRNA-seq data into a lower-dimensional latent space and analyze attention coefficients.