Including network knowledge into Cox regression models for biomarker signature discovery

Area of Science:

• Biomedical Informatics
• Computational Biology
• Genomics

Background:

• Biomarker gene signatures are crucial for personalized medicine but often lack reproducibility and interpretability.
• Integrating molecular interaction networks is a promising approach to enhance gene signature development.
• Predicting patient event risk using gene signatures is less explored than classification tasks.

Purpose of the Study:

• To investigate and compare eight network-informed methods for multivariable Cox proportional hazard models in breast cancer risk prediction.
• To evaluate the prediction performance, stability, and interpretability of gene signatures derived from these methods.
• To identify a robust and interpretable method for developing prognostic gene signatures.

Main Methods:

• Eight network integration methods were applied to multivariable Cox proportional hazard models.
• Cross-validation and multi-dataset comparisons were used to assess prediction performance.
• GeneRank-based filtering was explored as a technique for network information integration.

Main Results:

• GeneRank-based filtering demonstrated a simple, computationally efficient, and highly predictive approach.
• Network-integrated gene signatures showed improved reproducibility compared to traditional methods.
• The study identified specific network-informed methods that enhance risk prediction accuracy in breast cancer.

Conclusions:

• GeneRank-based filtering is a superior method for integrating network information into event time prediction models.
• This approach significantly improves the reproducibility and predictive power of gene signatures for breast cancer.
• The findings support the use of network-integrated biomarkers for more reliable clinical applications.