Bayesian variable selection in multinomial probit models to identify molecular signatures of disease stage

Area of Science:

• Statistics
• Bioinformatics
• Genomics

Background:

• Discrimination problems with high-dimensional data (more predictors than samples) are common in biological studies.
• Multinomial probit models are suitable for categorical outcomes but face challenges with many predictors.
• Identifying distinct molecular signatures is crucial for understanding disease progression.

Purpose of the Study:

• To develop a Bayesian variable selection approach for multinomial probit models in high-dimensional settings.
• To identify specific gene expression patterns that differentiate between two stages of rheumatoid arthritis.
• To provide a robust statistical framework for analyzing complex biological data.

Main Methods:

• Utilizing mixture priors and Markov chain Monte Carlo (MCMC) techniques for variable selection.
• Applying a Bayesian approach to multinomial probit models.
• Analyzing functional genomics data, specifically gene expression profiling.

Main Results:

• Successfully selected relevant variables that differ across classes in the context of rheumatoid arthritis.
• Identified potential molecular signatures associated with distinct disease stages.
• Demonstrated the efficacy of the proposed Bayesian method in a high-dimensional functional genomics problem.

Conclusions:

• The proposed Bayesian variable selection method is effective for high-dimensional discrimination problems.
• The identified molecular signatures can aid in characterizing different stages of rheumatoid arthritis.
• This approach offers a valuable tool for functional genomics and disease subtyping.