MicrobeWise: A Machine learning Toolkit for Predicting Syntrophic and Competitive Microbial Relationships

Microbial interactions are key determinants of community structure and ecosystem function, yet predicting interaction outcomes from genomic information remains challenging. In this study, we investigated whether pairwise microbial interactions can be inferred from genomic functional composition using genome-scale metabolic modeling combined with machine learning approaches. Genome-scale metabolic models from the AGORA2 collection were curated and linked to their corresponding genome assemblies, resulting in a non-redundant dataset of 1,378 microbial species. Pairwise metabolic interactions were simulated under different environmental conditions using the Cooperative tradeoff framework. Interaction outcomes were inferred from growth changes relative to single-species simulations and categorized based on normalized growth ratios. Genomic functional features were derived from KEGG Ortholog annotations and encoded as pairwise feature vectors capturing shared and asymmetric functional capabilities between microbial genomes. Supervised classification models, including a multilayer perceptron neural network and a Random Forest classifier, were trained separately for each environmental configuration to evaluate whether genomic composition alone can approximate interaction outcomes generated by metabolic simulations. The results indicate that genomic functional potential contains predictive signals for interaction outcomes, while environmental conditions strongly influence which interactions are realized.