Variational Bayes for high-dimensional proportional hazards models with applications within gene expression

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Analyzing high-dimensional sparse survival data presents challenges for Bayesian methods regarding scalability, variable selection, and uncertainty quantification.
• Existing Bayesian approaches often compromise uncertainty quantification or incur high computational costs.

Purpose of the Study:

• To develop an interpretable and scalable Bayesian proportional hazards model for prediction and variable selection in high-dimensional sparse survival data.
• To address the limitations of existing methods by providing both efficient computation and robust uncertainty quantification.

Main Methods:

• Developed a novel sparse variational Bayes method based on mean-field variational approximation.
• This approach overcomes the computational expense of Markov chain Monte Carlo.
• Offers a posterior distribution for parameters and natural variable selection via posterior inclusion probabilities.

Main Results:

• Extensive simulations demonstrate comparable or superior performance against state-of-the-art Bayesian variable selection methods.
• The method was successfully applied to two transcriptomic datasets with censored survival outcomes for variable selection.
• Uncertainty quantification provided interpretable patient risk assessments.

Conclusions:

• Sparse variational Bayes offers a scalable and interpretable solution for high-dimensional sparse survival data analysis.
• The method effectively balances variable selection, effect estimation, and uncertainty quantification.
• Freely available R package `survival.svb` facilitates practical application.