omicsGAT

omicsGAT applies a Graph Attention Network to integrate RNA-seq gene expression and sample relationships for cancer subtype analysis and related phenotype prediction.

Key Features:

• Graph Attention Network (GAT): Combines graph-based learning with an attention mechanism to assign varying attention coefficients to neighboring nodes.
• Multi-head Attention Mechanism: Employs multi-head attention to capture different importance levels from each neighbor of a sample.
• Integration with RNA-seq Data: Processes bulk RNA-seq and single-cell RNA-seq gene expression data to learn hidden representations.
• Graph Construction and Neighborhood Integration: Constructs graphs where nodes represent samples (patients or cells) and edges denote relationships based on gene expression, integrating neighborhood information into node representations.
• Embedding-based Prediction and Clustering: Produces embedding vectors that support disease phenotype prediction, patient stratification, and cell clustering.
• Adjacency Representation Comparison: Generates neighborhood representations that can provide more informative insights than traditional sample correlation-based adjacency matrices.

Scientific Applications:

• Cancer Subtype Analysis: Evaluates the importance of neighboring samples in networks to characterize and distinguish cancer subtypes.
• Patient Stratification and Cell Clustering: Enhances stratification of patients and clustering of cells to identify distinct cancer phenotypes.
• Phenotype and Outcome Prediction: Improves disease phenotype prediction and cancer outcome–related analyses using embedding vectors.
• Applied Datasets: Validated on The Cancer Genome Atlas (TCGA) bulk RNA-seq datasets including breast cancer and bladder cancer, and on two single-cell RNA-seq datasets.

Methodology:

Constructs a sample graph from gene expression where nodes are samples and edges reflect expression-based relationships, applies a Graph Attention Network with multi-head attention to compute attention coefficients over neighbors, and integrates neighborhood-weighted information into embedding vectors representing each sample.