Doe to time intensity and cost, the biological effects microbiome metabolites remain largely uncharacterized due to low-throughput and untargeted experimental approaches. “Prospective experimental validation confirmed the accuracy of our models and uncovered previously unknown effects of several metabolites.” according to a March 07 (updated April 17) 2026 Cell.com journal, iScience article.

The machine learning platform trained on publicly available drug development data to rapidly predict a wide array of chemical and biological properties of microbiome metabolites. The study was a joint collaboration between three departments of Duke University – Department of Biomedical Engineering, Computational Biology and Bioinformatics Program, and the Duke Microbiome Center.

Though some certain types of drugs or metabolites are chemically distinct from the other groups, the researchers observed a number of similar property distributions between drugs and microbiome metabolites. They were able to identify and utilize the machine learning as indicative benchmarks for properties crucial to a molecule’s bioavailability and their “drug-likeness.” The analysis suggests that microbiome metabolites share similar molecular patterns and properties with drug-like structures, which indicates that machine learning models trained on drug-like structures may be well suited for guided elucidation of the properties and biological effects of microbiome metabolites, according to the Duke study.

https://www.cell.com/iscience/fulltext/S2589-0042(26)00657-7