gCoda: Conditional Dependence Network Inference for Compositional Data.

Area of Science:

• Microbiome research
• Computational biology
• Network inference

Background:

• High-throughput sequencing of 16S rRNA genes allows direct analysis of natural microbial communities.
• Understanding direct microbial interactions is key to community assembly and maintenance.
• Compositionality and high dimensionality are major challenges in microbiome data analysis.

Purpose of the Study:

• To develop a robust method for inferring direct microbial interactions from compositional and high-dimensional microbiome data.
• To address the challenges of compositionality and dimensionality in microbiome network analysis.
• To propose a novel penalized maximum likelihood method for estimating sparse inverse covariance structures.

Main Methods:

• Modeling microbiome data using the logistic normal distribution to handle compositionality.
• Inferring direct interactions via conditional dependence networks under the logistic normal assumption.
• Developing gCoda, a penalized maximum likelihood method with a Majorization-Minimization algorithm to estimate sparse inverse covariance for high-dimensional data.

Main Results:

• gCoda effectively models the compositional nature of microbiome data.
• The proposed Majorization-Minimization algorithm efficiently solves the gCoda optimization problem.
• Simulation studies demonstrate gCoda's superior performance in edge recovery compared to existing methods like SPIEC-EASI.
• gCoda shows improved accuracy in inferring direct microbial interactions from mouse skin microbiome data.

Conclusions:

• gCoda provides a powerful and accurate approach for microbial interaction network inference.
• The method successfully addresses key challenges in microbiome data analysis, namely compositionality and high dimensionality.
• gCoda represents a significant advancement for microbiome research, enabling deeper understanding of microbial communities.