Inference for microbe-metabolite association networks using a latent graph model.

Area of Science:

• Microbiology
• Bioinformatics
• Network Science

Background:

• Correlation networks are widely used to infer microbe-metabolite associations.
• Existing methods for false discovery rate (FDR) control often assume weak dependence, limiting power for complex networks.

Purpose of the Study:

• To develop a novel inference procedure for detecting microbe-metabolite associations with enhanced power and FDR control.
• To incorporate latent network structures for improved association inference.

Main Methods:

• Proposed a novel inference procedure using a bipartite stochastic block model to capture latent network structures.
• Developed a variational expectation-maximization (EM) algorithm for parameter estimation and graph learning.
• Integrated the learned graph into the statistical testing procedure for association detection.

Main Results:

• The proposed method demonstrates enhanced power in detecting significant microbe-metabolite associations compared to existing methods.
• The procedure effectively controls the false discovery rate (FDR).
• The method provides module clustering of microbes and metabolites, aiding biological interpretation.

Conclusions:

• The novel inference procedure offers a powerful and reliable approach for analyzing microbe-metabolite association networks.
• Modeling latent network structures improves the detection of significant associations and network topology.
• The method is applicable to biological datasets, such as the analysis of bacterial vaginosis.