Dynamic Bernstein GCN for Pan-Cancer Subtype Classification Using RNA-Seq and CNV Data.

Area of Science:

• Computational biology and bioinformatics
• Machine learning in oncology
• Systems biology and network analysis

Background:

• Cancer subtype classification is challenging due to complex multi-omics interactions.
• Conventional machine learning models struggle to represent these interactions effectively.
• Graph Convolutional Networks (GCNs) use biological topologies but have fixed propagation, limiting adaptability.

Purpose of the Study:

• To introduce a novel architecture, the Dynamic Bernstein Graph Convolutional Network (DB-GCN), for cancer subtype classification.
• To enable topology-aware learning using adaptive spectral propagation with Bernstein polynomials.
• To support both single-omics and multi-omics data integration within a graph framework.

Main Methods:

• Developed DB-GCN with adaptive spectral propagation using Bernstein polynomials, avoiding eigendecomposition.
• Integrated single-omics (RNA) and multi-omics (RNA+CNV) data.
• Employed a dual-stream design combining a Bernstein graph stream and an omics multilayer perceptron.

Main Results:

• Achieved high accuracy in pan-cancer subtype classification across 28 TCGA subtypes using multi-omics data (86.05% ± 0.83 on STRING).
• Identified putative biomarker genes (e.g., KLK11, OR4F15, UBE2DNL) using SHAP analysis.
• Found that 12 of the top 50 identified genes map to KEGG cancer pathways.

Conclusions:

• DB-GCN offers an accurate and interpretable graph-based framework for pan-cancer subtype classification.
• The model effectively captures complex gene interactions for improved precision oncology.
• DB-GCN facilitates robust biomarker discovery for cancer research.