A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer

Area of Science:

• Oncology
• Computational Biology
• Bioinformatics

Background:

• Predicting cancer treatment response is challenging due to diverse therapies.
• Current biomarkers lack robustness and interpretability, hindering personalized oncology.

Purpose of the Study:

• To develop a deep learning framework (SubNetDL) for predicting treatment response using mutation data and network propagation.
• To identify robust and interpretable biomarkers for patient stratification across various cancer types and treatments.

Main Methods:

• SubNetDL integrates subclonal mutation profiles with protein-protein interaction networks using network propagation.
• The framework leverages somatic mutations alone, enabling broad applicability across cancer types and treatment modalities.
• Network propagation enhances the analysis of mutation data within biological network contexts.

Main Results:

• SubNetDL demonstrated strong predictive performance across 10 TCGA cancer-drug combinations (median AUROC = 0.74).
• The framework generalized effectively to independent immunotherapy datasets (median AUROC = 0.77).
• SubNetDL identified treatment-specific candidate biomarker genes, including those not traditionally central in networks.

Conclusions:

• SubNetDL provides a robust and interpretable deep learning framework for predictive biomarker discovery in oncology.
• The approach facilitates patient stratification by integrating mutation profiles with network context.
• This method advances personalized medicine by offering a versatile tool for diverse cancer treatments.