EPIC: Event Prototyping via Information Constrained graph learning for personalized cancer driver gene prediction

Area of Science:

• Computational Biology
• Genomics
• Oncology

Background:

• Distinguishing patient-specific driver mutations from passenger alterations is crucial for precision oncology.
• Existing graph-based methods struggle to preserve genomic context, obscuring subtle driver signals.
• This impedes the identification of individualized oncogenic events for personalized cancer therapy.

Purpose of the Study:

• To introduce EPIC (Event Prototyping via Information Constrained Graph Learning), a novel framework for driver mutation prediction.
• To redefine driver prediction as a metric learning task in an event embedding space.
• To develop an information-constrained learning strategy that preserves low-frequency driver signals.

Main Methods:

• EPIC employs a metric learning approach in an event embedding space.
• An information-constrained learning strategy is used to prevent feature collapse and preserve low-frequency signals.
• The framework redefines driver prediction beyond traditional node-centric graph methods.

Main Results:

• EPIC significantly outperforms established baseline methods on large-scale cancer cohorts.
• The model successfully prioritizes low-frequency driver variants, identifying mechanisms of drug resistance and metastasis.
• Clinical actionability analysis shows EPIC expands eligibility for targeted therapies.

Conclusions:

• EPIC offers a robust, context-aware solution for personalized cancer driver discovery.
• The framework bridges the gap between genomic data and actionable therapeutic insights.
• EPIC enhances the potential of precision oncology by improving driver mutation identification.