CGMega: explainable graph neural network framework with attention mechanisms for cancer gene module dissection.

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Cancer arises from complex interactions among multiple genes, forming gene modules, rather than single gene abnormalities.
• Understanding these gene modules is crucial for deciphering cancer development and heterogeneity.

Purpose of the Study:

• To develop and validate CGMega, an explainable, graph attention-based deep learning framework for cancer gene module dissection.
• To integrate multi-omics information for enhanced cancer gene prediction and mechanistic insights.

Main Methods:

• Leveraging model-agnostic interpretation approaches.
• Developing CGMega, a deep learning framework utilizing graph attention mechanisms.
• Applying CGMega to breast cancer cell line and acute myeloid leukemia (AML) patient data.

Main Results:

• CGMega outperforms existing methods in cancer gene prediction.
• Identified a high-order gene module involving the ErbB family and tumor factors NRG1, PPM1A, and DLG2.
• Discovered 396 candidate AML genes, with enrichment of known or candidate genes within specific modules.
• Identified patient-specific AML genes and their associated modules, revealing cancer heterogeneity.

Conclusions:

• CGMega is a powerful tool for dissecting cancer gene modules.
• The framework provides high-order mechanistic insights into cancer development and heterogeneity.
• CGMega facilitates the integration of multi-omics data for comprehensive cancer analysis.