engGNN: A Dual-Graph Neural Network for Omics-Based Disease Classification and Feature Selection.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Omics data (transcriptomics, proteomics, metabolomics) are crucial for understanding disease but present challenges like high dimensionality and small sample sizes.
• Existing Graph Neural Network (GNN) methods for omics data analysis often rely on either external or data-driven graphs, limiting their ability to capture comprehensive information.

Purpose of the Study:

• To develop a novel dual-graph framework, engGNN, that integrates both external biological networks and data-driven graphs for improved omics data analysis.
• To enhance the predictive performance and interpretability of GNNs in disease classification and biomarker discovery using high-dimensional omics data.

Main Methods:

• The engGNN framework constructs a biologically informed undirected feature graph from established network databases.
• It complements the undirected graph with a directed feature graph derived from tree-ensemble models, creating a dual-graph approach.
• This dual-graph design generates more comprehensive embeddings for omics data.

Main Results:

• engGNN demonstrated superior performance compared to state-of-the-art baselines in extensive simulations and real-world gene expression data analysis.
• The framework achieved improved predictive accuracy in disease classification tasks.
• engGNN provided interpretable feature importance scores, facilitating biologically meaningful discoveries like pathway enrichment analysis.

Conclusions:

• engGNN offers a robust, flexible, and interpretable framework for analyzing high-dimensional omics data.
• The dual-graph approach effectively addresses the limitations of existing GNN methods in omics research.
• This framework holds significant potential for advancing disease classification and biomarker discovery.