A Machine Learning Approach to logBB Prediction Combining Molecular Docking and Quantum-Chemical Descriptors

Blood–brain barrier (BBB) permeability is a major challenge in the development of drugs targeting the central nervous system (CNS). Computational prediction methods can help prioritize candidate compounds and reduce the experimental workload during early drug discovery. In this study, machine learning models were developed to predict the blood-to-brain concentration ratio expressed as its decimal logarithm (logBB). The objective was to investigate whether molecular docking scores obtained with AutoDock Vina for P-glycoprotein (P-gp) and breast cancer resistance protein (BCRP), together with quantum-chemical descriptors, could provide additional information for logBB prediction. The dataset comprised 685 compounds. A linear regression analysis identified eight factors associated with logBB: hydrogen bond donors, gas-phase electronic energy, electronic energy difference between water and hexane, number of heavy atoms, LogP, gas-phase geometry optimization time, number of aromatic rings, and fraction of sp3-hybridized carbon atoms. Notably, the electronic energy difference between water and hexane showed a stronger association with logBB than the corresponding Gibbs free energy difference, suggesting that electronic energy descriptors may provide complementary information. Higher-level quantum-chemical calculations will be required to investigate this finding further. Twenty-one machine learning models were evaluated. The selected support vector regression model using a radial basis function kernel (SVR-RBF; C = 0.869, ε = 0.135, γ = 0.165) was trained on 585 compounds and evaluated on a held-out validation set of 100 compounds, achieving an R2 of 0.605, an MAE of 0.368, and an RMSE of 0.459. Finally, the computational time analysis showed that descriptor generation was strongly dependent on molecular weight, with quantum-chemical calculations and molecular docking accounting for most of the computational cost. Overall, this work highlights the potential of combining physicochemical, molecular docking, and quantum-chemical descriptors for machine learning-based logBB prediction.