Bayesian compositional regression with structured priors for microbiome feature selection.

Area of Science:

• Microbiome Research
• Statistical Bioinformatics
• Human Health and Disease

Background:

• The human microbiome is crucial for health and disease, yet standard analysis methods struggle with its unique data characteristics.
• Microbiome data is compositional (fixed-sum constraint) and features (operational taxonomic units) often represent functionally similar microbes.

Purpose of the Study:

• To develop an advanced variable selection technique specifically designed for the complexities of microbiome data.
• To improve the identification of microbial features associated with clinical outcomes, such as body mass index.

Main Methods:

• Proposed a novel variable selection method incorporating a generalized transformation and z-prior to address compositional data.
• Utilized an Ising prior to promote the selection of genetically similar and functionally related microbiome features.
• Evaluated the method's performance against existing penalized approaches using simulations and real-world gut microbiome data.

Main Results:

• The proposed method demonstrated superior performance compared to current penalized variable selection techniques.
• Successfully identified key gut microbiome features related to body mass index in real-world data analysis.

Conclusions:

• The novel statistical approach effectively handles the compositional nature and feature relatedness inherent in microbiome data.
• This method offers a more robust framework for linking microbiome composition to human health indicators, advancing personalized medicine.