Machine learning-guided design of human gut microbiome dynamics in response to dietary fibers

Abstract

Dietary fibers are key modulators of human gut microbiome dynamics and functions, yet we lack the ability to predict community dynamics and functions in response to fibers. We integrate machine learning, Bayesian optimization, and high-throughput community construction to investigate how dietary fibers shape health-relevant functions of human gut microbial communities. To efficiently navigate the landscape of fiber-microbiome interactions, we implemented a design-test-learn cycle to identify fiber-species combinations that maximize a multi-objective function capturing beneficial community properties. Our model-guided approach revealed a highly butyrogenic and robust ecological motif characterized by the copresence of inulin, Bacteroides uniformis, and Anaerostipes caccae and a higher-order interaction with Prevotella copri. In germ-free mice, model-designed species-fiber combinations impacted colonization, community diversity and short-chain fatty acid profiles. In sum, our work establishes a novel framework for designing microbial communities with desired functions in response to key nutrients such as dietary fibers.