Improving the generalizability of peptide permeability prediction models using physics-based ensemble energetics

Peptides are an increasingly attractive therapeutic modality, combining the precision of antibodies with the drug-like properties and scalability of small molecules. Their richer conformational landscape and the lack of available training data when compared to small molecules, however, make them difficult to model with conventional machine learning (ML) approaches. In particular, very few accurate and generalizable models exist for predicting membrane permeability, a crucial drug-like property for clinical success. Here, we introduce PepPerm3D, a generalizable ML model for permeability prediction augmented by large-scale conformational sampling and physics-based quantum mechanical calculations. While PepPerm3D is competitive with state-of-the-art models under random and scaffold-based splits, our aim was to critically assess its predictive ability under a publication split to better mimic the conditions of a true drug discovery campaign. By combining publication-normalized targets with explicit thermodynamic descriptors of conformer energetics, basin populations, polarity shielding, intramolecular hydrogen bonding, and solvent reorganization, PepPerm3D improved ranking across unseen publications and increased top-versus-bottom-quartile permeability discrimination from an AUC of 0.67 to 0.75. We show specific examples of congeneric peptide series where 3D physics features greatly improve predictivity and provide an atomic-level rationalization of the improvement from the sampled conformational ensembles.