Sparse Bayesian Group Factor Model for Feature Interactions in Multiple Count Tables Data.

Area of Science:

• Microbiome research
• Statistical modeling
• Bioinformatics

Background:

• Microbiome studies generate complex, high-dimensional count data across multiple domains.
• Existing models struggle to capture intricate interactions between microorganisms in these datasets.
• Next-generation sequencing data presents challenges due to large variability and excess zeros.

Purpose of the Study:

• To develop a novel sparse Bayesian group factor model (Sp-BGFM) for analyzing multiple microbiome count tables.
• To effectively capture inter-domain microorganism interactions.
• To incorporate covariate effects on microbial abundances.

Main Methods:

• Developed a sparse Bayesian group factor model (Sp-BGFM) utilizing a rounded kernel mixture model with a Dirichlet process (DP) prior.
• Employed log-normal mixture kernels for count vectors and a group factor model for the covariance matrix.
• Introduced a Dirichlet-Horseshoe (Dir-HS) shrinkage prior for joint sparsity in factor loading vectors and incorporated regression for covariate effects.

Main Results:

• The Sp-BGFM demonstrated superior performance in high-dimensional applications due to joint sparsity induced by the Dir-HS prior.
• The flexible DP model effectively handled large variability and excess zeros in observed counts.
• Robust estimation of microorganism interactions and covariate effects was achieved.

Conclusions:

• Joint sparsity is crucial for accurate analysis of high-dimensional microbiome data.
• The proposed Sp-BGFM provides a flexible and robust framework for microbiome count data analysis.
• The model offers significant benefits for understanding microbial community structures and interactions.