A hybrid framework for disease biomarker discovery in microbiome research combining Bayesian networks, machine learning, and network-based methods.

Microbiome research faces two central challenges, namely constructing reliable networks, where nodes represent microbial taxa and edges represent their associations, and identifying significant disease-associated taxa. To address the first challenge, we developed CMIMN, a novel R package that applies a Bayesian network framework based on conditional mutual information to infer microbial interaction networks. To further enhance reliability, we construct a consensus microbiome network by integrating results from CMIMN and three widely used methods, including Sparse Inverse Covariance Estimation for Ecological Association Inference (SPIEC-EASI), Semi-Parametric Rank-based correlation and partial correlation Estimation (SPRING), and Sparse Correlations for Compositional Data (SPARCC). This consensus approach, which overlays and weights edges shared across methods, reduces inconsistencies and provides a more biologically meaningful view of microbial relationships. To address the second challenge, we designed a multi-method feature selection framework that combines machine learning with network-based strategies. Our machine learning pipeline applies distinct algorithms and identifies key taxa based on their consistent importance across models. Complementing this, we employ two network-based strategies that prioritize taxa based on centrality differences between networks constructed from healthy samples and disease-affected samples, as well as a composite scoring system that ranks nodes using integrated network metrics. We applied CMIMN on soil microbiome data from potato fields affected by common scab disease. Bootstrap analysis confirmed the robustness of CMIMN, and the consensus network further improved stability and interpretability. The multi-method framework enhances confidence in identifying soil microbial taxa associated with potato disease. Notably, we identified