Profiling Human Microbiome Metabolites via Computational and Experimental Approaches

Introduction:: The human microbiome contributes to various physiological processes, including digestion, the development of chronic inflammatory diseases, and modulating immunotherapy responses. Such host-microbiome interactions are mediated in part by the vast numbers of metabolites produced by the microbiome that exert protein modulatory effects in their host. However, current experimental approaches in microbiome research are low-throughput, prohibiting systematic testing of all host-microbiome interactions. We aim to accelerate this work by creating machine learning models that predict pharmacokinetic properties and biological effects of microbiome metabolites. Such predictions will facilitate the study of microbiome-host interactions by shifting from untargeted screens to targeted validation experiments. Our research will provide a machine learning platform to predict biological profiles of microbiome metabolites to accelerate the discovery of metabolites relevant across clinical fields as biomarkers or as potential therapeutic interventions.


Materials and Methods:: We have generated the world’s largest database of human microbiome metabolites by curating data from HMDB, VMH.life, KNAPSACK^, and Han’s group. This dataset contains 5,072 microbiome metabolites produced by more than 2000 different commensal gut microbe strains with canonical SMILES structures. The pharmacokinetic properties and biological effects of metabolites are predicted by machine learning models and compared to FDA-approved drugs. Both regression and classification models are used depending on the given datasets and predicting properties. Training datasets are obtained from Therapeutics Data Commons, which provides datasets of a wide range of drug properties. We use various standard machine learning models including Random Forest, Gradient Boosting, Support Vector Machine and deep models such as Multilayer Perceptron or Recurrent Neural Networks to predict pharmacokinetic properties of microbiome metabolites. The best performing models for each property are determined by stratified or scaffold-based cross-validations. The top predicted metabolites in each category are validated using external datasets composed of human metabolites or performing in vitro experiments. For example, the “human intestine absorption” model can be validated by using human plasma metabolites derived from the gut microbiome. Since these plasma metabolites need to be absorbed through the intestinal barrier first, we can consider them as “positive”. Accuracy is then calculated to determine the success of the machine learning model. Drug-induced liver injury (DILI) was validated by performing cell viability assays of HEPG2 liver cells after exposure to the metabolites. A confusion matrix is created based on the cell assay results and accuracy/balanced accuracy is calculated.


Results, Conclusions, and Discussions:: Microbiome metabolites overlap in chemical space with FDA-approved drugs, while there is also a large space that is only covered by metabolites, suggesting possible interest of drug discovery. 39.2% of metabolites have “drug-like properties” according to Lipinski's “rule of five.” We found that metabolites and drugs share similar property distributions in multiple relevant properties such as logP and molecular weight. Next, we developed regression models to predict toxicity, solubility, plasma protein binding, and cell permeability for the microbial metabolites and compare their predicted property values to FDA-approved drugs. Metabolites and drugs show similar distributions compared to drugs in predicted toxicity and solubility but are predicted to have higher plasma protein binding compared to drugs, indicating a more efficient diffusion in the bloodstream. Classification models were developed to predict human intestinal absorption, blood brain barrier (BBB) penetration, DILI, and bioavailability. More than half (53.4%) of microbiome metabolites are predicted to be absorbed through the host’s intestinal barrier, supporting a large potential for systemic crosstalk of metabolites and host physiology beyond the gastrointestinal system. Human intestinal absorption and BBB penetration predictions were validated using external datasets showing 92% and 86% accuracy, respectively. For DILI validation, 7 predicted positive and 8 predicted negative metabolites were tested in HEPG2 cytotoxicity assays and provided 78.6% balanced accuracy. TD₅₀ (Toxic Dose) of positive metabolites were comparable to drugs with known liver toxicity. In summary, we developed and validated a database and machine learning platform to predict pharmacokinetic properties and biological effects of human microbiome metabolites, compared these properties to FDA-approved drugs, and validated several of our predictions. The provided methodology and findings will facilitate the study of microbiome-host interactions and accelerate the discovery of metabolites in clinical applications as biomarkers and therapeutic adjuvants.




Acknowledgements (Optional): :

References (Optional): :