Mechanistic and machine learning models design human gut consortia that robustly inhibit Clostridioides difficile

Abstract

Identifying design principles for robust inhibition of human pathogens is a major goal of microbiome engineering. By building synthetic microbial communities from the bottom-up guided by Bayesian active learning, we investigate the Clostridioides difficile growth landscape across thousands of species-metabolite conditions. Mechanistic consumer resource and machine learning models uncover significant interactions linking metabolites and species, and exhibit concordance with microbial interactions identified in human microbiome datasets. Guided by machine learning and mechanistic models, we elucidate microbial communities capable of robustly inhibiting C. difficile across diverse nutrient environments in vitro. Metabolomic profiling identified sorbitol, mannitol, and proline as key metabolites mediating community-driven C. difficile inhibition. Model-optimized communities significantly reduced C. difficile colonization in the murine gut whereas a model-designed non-inhibitory community did not exhibit this property. Communities display consistent metabolomic patterns in vitro and in vivo, suggesting that resource competition is a design principle of robust inhibition. Together, these findings establish an integrated experimental and computational framework for the rational design of microbial communities that robustly suppress pathogens across diverse environmental contexts.