Learning from SILCS FragMaps: Combining Physics-Based Energies with Artificial Intelligence for Rapid Drug Permeability Prediction

Prediction of drug passive permeability represents an important component of rational drug design, requiring continued development of methods to address this problem. Artificial intelligence (AI) methods have long been shown to be of utility for prediction of compound permeability, though they are limited by the range of chemical space in the data sets used in the training of the models. To address this, in the present study, the site identification of ligand competitive saturation (SILCS) methodology is used to calculate free energy profiles for compounds across 3 model lipid bilayers: a POPC–cholesterol bilayer, a bilayer model mimicking parallel artificial membrane permeability assay (PAMPA) experiments, and a bilayer modeling the blood–brain barrier. We present a rapid workflow generating around 15,000 such profiles that accounts for the passive permeation of more than 2500 unique ligands for 3 different membranes. The free energy profiles, which account for all accessible protonation states of the compounds, are then used as features in a collection of AI models. As the physics-based free energy profiles are calculated based on the functional group types in the compounds combined with the SILCS FragMaps calculated for the bilayers, their use potentially improves model transferability by overcoming the limitation of the applicability of the AI models to the chemical space included in the training set, assuming the SILCS-based energies are representative of barriers to compound permeabilities. The resulting models are shown to be almost comparable with available passive permeability models reported in the literature while additionally offering energetic information about the contribution of functional groups to the permeation barriers that may be used for rational compound optimization. Limitations in the cell-based training data, including the presence of drug transporters, are discussed.