An Interpretable Deep Learning Framework for Biomarker Discovery in Complex Disease Survival Outcomes

Area of Science:

• Biomedical data science
• Computational biology
• Genomics and bioinformatics

Background:

• Identifying biomarkers for complex disease survival is crucial for understanding disease mechanisms and advancing precision medicine.
• Time-to-event data presents challenges due to its complexity, including non-linear interactions and high dimensionality.
• Conventional survival data modeling approaches struggle with these complexities.

Purpose of the Study:

• To propose SurvDNN, a deep neural network framework tailored for survival outcome modeling.
• To enhance model robustness and mitigate overfitting in survival data analysis.
• To enable interpretable biomarker discovery through an extended Permutation-based Feature Importance Test (PermFIT).

Main Methods:

• Developed SurvDNN, a deep neural network framework for survival outcomes.
• Incorporated bootstrapping-based regularization to reduce overfitting.
• Implemented a stability-driven filtering algorithm for improved model robustness.
• Extended Permutation-based Feature Importance Test (PermFIT) for interpretable biomarker quantification.

Main Results:

• SurvDNN demonstrated superior performance compared to existing machine learning methods in simulations and real-world data.
• Achieved higher accuracy in both biomarker identification and predictive modeling.
• PermFIT provided robust quantification of individual biomarker contributions.

Conclusions:

• SurvDNN coupled with PermFIT offers a powerful, interpretable, and robust tool for biomarker-driven survival modeling.
• The framework holds significant potential for advancing precision medicine in complex diseases like cancer and cardiovascular disorders.
• An open-source R package for SurvDNN is available for public use.