Learning a mixture of microbial networks using minorization-maximization.

Area of Science:

• Microbiome research
• Computational biology
• Network inference

Background:

• Microbial community interactions are crucial for community behavior and member abundance.
• Microbial networks model these interactions using taxa as nodes and pairwise interactions as edges.
• Existing methods assume a single network per sample-taxa matrix, overlooking factor-driven network variations.

Purpose of the Study:

• To develop a computational framework for inferring multiple microbial networks from a single sample-taxa matrix.
• To address the limitation of existing methods that assume only one network per dataset.
• To model how environmental or host factors can lead to distinct microbial interaction networks within a single study.

Main Methods:

• A mixture model framework (MixMPLN) is proposed, utilizing a mixture of K Multivariate Poisson Log-Normal distributions.
• Parameter estimation employs a maximum likelihood framework with minorization-maximization, gradient ascent, and block updates.
• The approach was validated using synthetic datasets (absolute, compositional, and normalized counts) and incorporated an l1-penalty for sparse network recovery.

Main Results:

• The developed mixture model successfully infers multiple microbial networks from sample-taxa data.
• Performance was assessed on various synthetic data types, demonstrating the framework's robustness.
• The method effectively recovers sparse networks, providing insights into complex microbial associations.

Conclusions:

• The proposed mixture model framework advances microbial network inference by accommodating multiple networks.
• This approach provides a more realistic representation of microbial communities influenced by multiple factors.
• MixMPLN offers a valuable tool for analyzing complex microbiome data and understanding context-specific microbial interactions.