Prediction of human intestinal passive membrane permeability using physiologically informed machine learning models

Predicting human intestinal passive membrane permeability (P_eff) is essential for reliable application of physiologically based biopharmaceutics models (PBBMs). Although in vitro cell-based assays and animal studies are widely used, their predictive performance is often limited by inter-laboratory variability and species differences. Furthermore, accurate prediction of intrinsic permeability requires consideration of physiological factors, particularly correction for the unstirred water layer (UWL) and exclusion of transporter-mediated transport. We developed machine learning models to predict human jejunal P_eff directly from chemical structure using a physiologically curated dataset derived from in vivo human jejunal permeability data. The dataset was refined by excluding transporter substrates and correcting for UWL resistance to estimate intrinsic cellular permeability (P_cell). To maximize predictive performance while maintaining robustness with a limited dataset, two feature selection strategies were evaluated: a data-driven approach to identify novel structure-permeability relationships and a hypothesis-driven approach based on established physicochemical principles. The practical applicability of the developed models was assessed by incorporating the predicted Peff values into a PBBM. The data-driven model demonstrated superior generalization performance and accurately reproduced observed human plasma concentration-time profiles. These findings demonstrate that the proposed approach can robustly predict intrinsic human intestinal membrane permeability from chemical structure alone.