Pathway aggregation for survival prediction via multiple kernel learning.

Area of Science:

• Genomics
• Bioinformatics
• Cancer Research

Background:

• Predicting cancer prognosis from high-dimensional genomic data is challenging due to numerous features and complex relationships.
• Integrating biological pathway information can reduce dimensionality and enhance prediction accuracy and interpretability.

Purpose of the Study:

• To propose novel multiple kernel learning (MKL) methods for censored survival outcomes in cancer prognosis.
• To extend existing kernel machine frameworks to handle nonlinear pathway effects and improve prediction.

Main Methods:

• Developed MKL methods for a general survival modeling framework with convex objective functions.
• Applied the methods under Cox proportional hazards and semiparametric accelerated failure time models.
• Utilized grouping of genomic features into biological pathways and networks.

Main Results:

• Numerical studies showed the proposed MKL methods perform well in finite samples.
• The MKL approach demonstrated potential to outperform models with linear effects or without pathway knowledge.
• The methods were successfully applied to two cancer datasets.

Conclusions:

• The proposed MKL methods offer a robust and interpretable approach for cancer prognosis prediction using genomic data.
• Integrating biological pathway information via MKL can significantly improve prediction accuracy for survival outcomes.
• This work advances the application of machine learning in personalized cancer medicine.