Constructing Pathway-Based Priors within a Gaussian Mixture Model for Bayesian Regression and Classification.

Area of Science:

• Genomics
• Bioinformatics
• Statistical Learning

Background:

• Gene-expression data analysis is crucial for translational genomics, enabling classification and regression.
• Optimal learning requires known feature-label or predictor-target distributions, which are often unavailable in practice.
• Prior knowledge integration is essential for robust analysis, particularly with small sample sizes.

Purpose of the Study:

• To extend the Regularized Expected Mean Log-Likelihood Prior (REMLP) methodology to Gaussian mixture models (GMMs) for scenarios where sample labels are unknown.
• To develop a novel prior construction and update strategy for optimal Bayesian classification and regression.
• To enhance the performance of gene-expression-based analyses in translational genomics.

Main Methods:

• Extension of REMLP to a Gaussian mixture model (GMM) framework to handle unknown sample labels.
• Prior construction and update using Bayesian sampling to generate Monte Carlo approximations.
• Application to phenotype classification using prior knowledge of colon cancer pathways.

Main Results:

• The GMM REMLP prior demonstrates superior performance compared to the Expectation-Maximization (EM) algorithm for small datasets.
• Successful application in phenotype classification, leveraging pathway information for colon cancer.
• Provides accurate Monte Carlo approximations for optimal Bayesian regression functions and classifiers.

Conclusions:

• The GMM REMLP prior offers a significant advancement for gene-expression analysis when labels are unobserved.
• This method provides a robust approach for classification and regression in translational genomics, particularly beneficial for limited data.
• The integration of pathway information enhances the predictive power of Bayesian learning models.