A Mechanism-based Prediction Model of P-glycoprotein Compound Efflux
Yulong GOU *1, Suyong RE2, Kenji MIZUGUCHI1, Chioko NAGAO1
1Institute for Protein Research, The University of Osaka
2National Institutes of Biomedical Innovation, Health and Nutrition
P-glycoprotein (P-gp) is widely expressed in human cell membranes and contributes to expelling a broad range of substances from cells. Overexpression of P-gp in tumor cells leads to multidrug resistance (MDR) and affects the ADME properties of drugs. Predicting the efflux ratio of drugs helps improve the success rate of drug discovery but remains a challenge. Although some machine-learning-based prediction models have been developed, traditional machine learning methods are often highly data-dependent. Gunaydin et al. proposed a mechanism-based prediction model of the net efflux ratio (NER), where the difference in solvation free energy between water and chloroform was used to model the dissociation process of compounds from the P-gp binding site to the extracellular space. Previously, we validated this method using a dataset from Watanabe et al. and observed non-negligible outliers among highly hydrophilic non-substrates. Structural and physicochemical analysis revealed that these compounds had a higher number of hydrogen bond donors, suggesting that both intra- and intermolecular hydrogen bonding interactions are important in the P-gp-mediated efflux process.
Here, we developed a physicochemical model that extends Gunaydin’s approach by incorporating the binding process of compounds to P-gp, as well as the effect of intramolecular hydrogen bonding in the dissociation process. We performed 100 cyclic conformational searches for each compound using the MMFF94 force field in RDKit to model intramolecular interactions. We investigated the solvation free energy of these molecules in chloroform and water at the M062X/6-31G* level of IEFPCM theory, and the pKa values were corrected with the MolGpKa software. Since there was no available crystal structure of human P-gp in the inward-facing state, we generated a P-gp structure in this conformation using AlphaFold2.
For each compound, we generated 1,000 distinct docking poses using the AutoDock Smina program and then ranked them to identify the most stable binding modes and their corresponding binding affinities. Based on the known mechanism of P-gp-mediated efflux, we also hypothesized that molecular size affects the binding strength and the efflux process. Based on these insights, we developed a sigmoidal model linking the NER to three key physicochemical parameters: solvation free energy, binding affinity, and molecular size. This model reflects the saturable kinetics characteristic of biological transport systems.
To effectively integrate these parameters into a robust predictive model, we trained an XGBoost model using 5-fold cross-validation on our training dataset of 385 compounds and validated its performance on an independent test set of 48 compounds. Our model achieved an AUC-ROC of 0.89, with 94% accuracy for compounds with an NER < 2 and 90% recall for compounds with an NER ≥ 2.