Latent Network Estimation and Variable Selection for Compositional Data Via Variational EM.

Area of Science:

• Statistical modeling
• Bioinformatics
• Network analysis

Background:

• Simultaneous network estimation and variable selection are recent statistical challenges.
• Count data, especially compositional data with sum constraints, require specialized methods.
• Understanding microbe-microbe and microbe-covariate interactions is crucial for microbiome research.

Purpose of the Study:

• Develop a novel method for simultaneous network estimation and covariate association for compositional count data.
• Improve accuracy in network recovery and variable selection.
• Apply the method to analyze microbiome data and understand microbial interactions.

Main Methods:

• A hierarchical Bayesian model with latent layers.
• Spike-and-slab priors for edge and covariate selection.
• A novel variational inference scheme with an expectation-maximization step for efficient posterior inference.

Main Results:

• The proposed model demonstrates superior accuracy in network recovery compared to existing methods through simulations.
• The method effectively identifies significant network interactions and covariate associations.
• Successful application to real-world microbiome data.

Conclusions:

• The developed algorithm provides an efficient and accurate approach for simultaneous network and covariate inference in compositional count data.
• The method enhances the understanding of complex interactions within biological systems like the human microbiome.
• The publicly available Python implementation facilitates broader adoption and research.