Explainable AI predicts how plant compounds may cross intestinal barriers

Phytochemicals, bioactive compounds from plants like flavonoids and ginsenosides, often struggle to cross intestinal barriers, limiting their potential in functional foods and drugs. Traditional methods involve laborious cell-based experiments to measure absorption, a bottleneck slowing down screening of candidate ingredients. Researchers at Sejong University and Korea Institute of Science and Technology (KIST) developed explainable machine learning models to predict key intestinal absorption properties directly from phytochemicals' molecular structures, pre-experiment.

Focusing on apparent permeability, transepithelial electrical resistance, and efflux ratio—measured in the human intestinal Caco-2 cell model—the team computed 5,003 molecular descriptors from 83 structurally diverse phytochemicals. Using statistical feature selection, recursive feature elimination, and repeated cross-validation, they evaluated 10 regression algorithms for each indicator, finding the best performance from a stacking ensemble model combining CatBoost, LightGBM, and Gradient Boosting with a linear regression meta-model.

This ensemble achieved high predictive accuracy for apparent permeability and efflux ratio but moderate performance for TEER. Cross-validation reached coefficients of determination up to 0.96 for apparent permeability. The study also built a classification model predicting compounds by efflux behavior, achieving an area under the curve of 0.96 in cross-validation.

Explainable artificial intelligence techniques revealed electronic, topological, and branching-related descriptors as key determinants of absorption, providing chemists with concrete clues for designing better-absorbed ingredients. These explainable QSPR models serve as practical prescreening tools, allowing researchers to prioritize bioavailable compounds and reserve costly laboratory assays for the most promising ones.