Bayesian compositional regression with microbiome features via variational inference.

Area of Science:

• Microbiology
• Statistical Bioinformatics
• Human Health

Background:

• The human microbiome is crucial for health, but analyzing its compositional nature (relative abundances) presents challenges.
• Microbiome data is high-dimensional with components varying by orders of magnitude, often overlooked in analyses.
• Understanding microbiome associations with phenotypes like body mass index requires robust statistical methods.

Purpose of the Study:

• To develop a scalable Bayesian hierarchical linear log-contrast model for compositional microbiome data.
• To address the challenges of high dimensionality, scale differences, and constrained parameter spaces in microbiome analysis.
• To compare the proposed Bayesian method against state-of-the-art frequentist compositional data analysis techniques.

Main Methods:

• A Bayesian hierarchical linear log-contrast model estimated using mean field Monte-Carlo co-ordinate ascent variational inference (CAVI-MC).
• Novel priors were developed to handle scale differences and constrained parameter spaces of compositional covariates.
• A reversible jump Monte Carlo Markov chain was employed to estimate intractable marginal expectations.

Main Results:

• The proposed Bayesian method demonstrated favorable performance compared to existing frequentist compositional data analysis methods.
• The CAVI-MC approach effectively scales to high-dimensional microbiome datasets.
• The model was successfully applied to real-world data, exploring the gut microbiome's relationship with body mass index.

Conclusions:

• The developed Bayesian hierarchical model offers a powerful and scalable solution for analyzing compositional microbiome data.
• This method provides a significant advancement in understanding the complex interplay between the microbiome and human health indicators.
• The application to body mass index highlights the model's utility in real-world biomedical research.