A hybrid framework for disease biomarker discovery in microbiome research combining Bayesian networks, machine...

Area of Science:

• Microbiome research
• Network inference
• Machine learning applications in ecology

Background:

• Microbiome studies face challenges in constructing accurate microbial association networks.
• Identifying specific microbial taxa linked to diseases is crucial for understanding host-microbe interactions.
• Existing methods for network inference and feature selection have limitations in reliability and interpretability.

Purpose of the Study:

• To develop a robust R package (CMIMN) for inferring microbial interaction networks using a Bayesian framework.
• To create a consensus network approach by integrating multiple inference methods for enhanced reliability.
• To design a multi-method feature selection framework combining machine learning and network analysis for disease-associated taxa identification.

Main Methods:

• Developed the CMIMN R package utilizing conditional mutual information for Bayesian network inference.
• Constructed a consensus microbiome network by integrating CMIMN with SPIEC-EASI, SPRING, and SPARCC.
• Implemented a machine learning pipeline and network-based strategies (centrality differences, composite scoring) for feature selection.

Main Results:

• The CMIMN package demonstrated robustness through bootstrap analysis on soil microbiome data.
• The consensus network approach improved the stability and biological interpretability of microbial relationships.
• The multi-method framework successfully identified key soil microbial taxa associated with potato common scab disease, including specific phyla, classes, and orders.

Conclusions:

• CMIMN provides a reliable tool for microbial network inference.
• The consensus network and multi-method feature selection enhance confidence in identifying disease-associated microbes.
• This integrated approach advances the understanding of soil microbiome roles in plant diseases.