Bayesian compositional generalized linear mixed models for disease prediction using microbiome data.

Area of Science:

• Microbiome research
• Computational biology
• Statistical modeling

Background:

• Predictive modeling of microbiome data is crucial for understanding disease susceptibility.
• Current methods often assume sparsity, overlooking the impact of minor microbial taxa.
• Real-world data frequently exhibits both large and small effect sizes.

Purpose of the Study:

• To develop a novel statistical framework, Bayesian Compositional Generalized Linear Mixed Models (BCGLMM), for analyzing compositional microbiome data.
• To enhance predictive accuracy by accounting for both moderate and minor microbial effects.
• To improve the understanding of disease susceptibility linked to the microbiome.

Main Methods:

• Developed BCGLMM incorporating a structured regularized horseshoe prior for sparsity and phylogenetic collaboration.
• Utilized a random effect term with a variance-covariance matrix to capture sample-related minor effects.
• Employed Markov Chain Monte Carlo (MCMC) algorithms via rstan for model fitting.

Main Results:

• Extensive simulations demonstrated BCGLMM's superior prediction accuracy compared to existing methods.
• The model effectively identifies moderate taxa effects and the cumulative impact of minor taxa.
• BCGLMM successfully predicted inflammatory bowel disease (IBD) using American Gut Data.

Conclusions:

• BCGLMM offers a powerful and accurate approach for microbiome-based disease prediction.
• The method's ability to integrate diverse effect sizes enhances predictive modeling capabilities.
• This framework advances the analysis of compositional microbiome data for biomedical applications.