Identification of Gene Regulatory Networks Associated with Breast Cancer Patient Survival Using an Interpretable Deep...

Area of Science:

• Biomedical research
• Computational biology
• Cancer genomics

Background:

• Artificial neural networks show promise in biomedical research but face challenges in survival analysis.
• Optimizing models for both accuracy and biological interpretability is crucial for clinical utility.

Purpose of the Study:

• To develop a deep neural network (MaskedNet) for identifying genes and pathways associated with breast cancer patient survival.
• To enhance biological insights from survival analysis models.

Main Methods:

• Developed MaskedNet, a deep neural network trained on TCGA breast cancer data.
• Interpreted model outputs using SHapley Additive exPlanations (SHAP) to assign feature importance.
• Validated findings in an independent breast cancer clinical trial.

Main Results:

• MaskedNet demonstrated higher accuracy than traditional Cox regression.
• Identified IFNG and PIK3CA genes and associated pathways linked to overall survival.
• Found that higher IFNG SHAP values correlated with better survival, linked to M1 macrophages and T cell infiltration.

Conclusions:

• MaskedNet effectively integrates survival analysis with biological interpretability in cancer research.
• The IFNG pathway's association with immune microenvironment and survival was validated in an independent cohort.