Machine Learning Guided Discovery of Microbiome Metabolite Bioactivities

Abstract

The human microbiome produces thousands of small molecules that influence host physiology, immunity, and disease progression. However, only a small fraction of these microbiome-derived metabolites has been characterized for their biological effects because most current experimental methods are often slow, expensive, and often limited to correlation-based analyses. As a result, the mechanisms by which microbiome metabolites affect host processes remain largely unknown. This dissertation aims to address this gap by developing and validating a machine learning–based framework for systematic prediction and experimental testing of microbiome metabolite bioactivities.To achieve this goal, we curated a dataset of microbiome metabolites as our testing chemical library and assembled diverse training sets of drug-like molecules to develop predictive machine learning models for absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties, as well as immunomodulatory activity through interleukin-8 (IL-8) stimulation. External datasets and in vitro experiments were employed to validate model performance. The applicability of this platform to various biological questions was tested by further refining and applying the pipeline to other targets. First, histone deacetylase 3 (HDAC3) was used as a case study to examine how optimizing training data composition improves prediction accuracy for microbiome-like compounds. Second, a multi-objective modeling approach was applied to identify microbial metabolites that inhibit Toll-like receptor 9 (TLR9) signaling and suppress breast cancer cell invasion. Predicted candidates were tested using biochemical and cellular assays, molecular docking simulations, and cell-based invasion models. The results demonstrate that many microbiome metabolites possess drug-like features that make them suitable for structure-based property prediction. In the ADMET and IL-8 models, most microbiome metabolites were predicted to be liver-safe, although in vitro testing identified several hepatotoxic compounds and previously unrecognized IL-8–stimulating metabolites. In the ix HDAC3 study, the tailored subsampling-based training approach improved predictive accuracy for metabolites, and molecular docking simulations supported the predicted metabolite–HDAC3 interactions. In the TLR9 study, multi-objective models identified two promising metabolites that significantly reduced invasive activity in vitro. Model performance varied across different evaluation tasks, ranging from moderate accuracy on external datasets to high accuracy in prospective experiments. Given the small sample sizes and inherent variability, these results should be interpreted qualitatively, but they collectively illustrate the framework’s capacity to generate biologically meaningful hypotheses linking microbial chemistry to host phenotypes. Collectively, this work prototypes a scalable computational and experimental pipeline for discovering bioactive microbiome metabolites. By integrating machine learning with biological validation, the dissertation aims to provide a proof-of-concept for a new approach to microbiome research that focuses on computational predictions as hypothesis generators. The presented framework provides a foundation for identifying microbiome-derived metabolites with therapeutic or biomarker potential and contributes to a deeper understanding of how microbial chemistry influences human health and disease.