Computational hybrid approach of PBPK simulation and CGCNN models for human intravenous pharmacokinetics

In drug discovery, it is not sufficient for a lead compound to exhibit high binding affinity for the target protein alone. It is equally important to predict an optimal pharmacokinetic (PK) profile that reflects the physicochemical properties of the compound. In this study, we aimed to predict human PK by jointly employing physiologically based pharmacokinetic (PBPK) simulation and chemical graph convolutional neural networks (CGCNN) models. Using a hybrid approach that combines PBPK simulation with CGCNN models, we successfully constructed a model capable of predicting human area under the concentration-time curve (AUC) and maximum concentration (Cmax) in plasma after intravenous bolus administration using only chemical structure as input. Furthermore, by integrating matched molecular pair (MMP) with fingerprint descriptors, we added interpretability to prediction models, enabling a visual understanding of the relationship between structural modifications and PK properties.