Developing a Quantum Physics-Inspired Machine Learning Model for Partition Coefficient Prediction and Comparing Its Performance with Empirical Descriptors and Molecular Graph-Based Models

Abstract

Predicting the partition coefficient (logP) is critical in pharmaceutical research for assessing the bioavailability of small molecules. We present a novel quantum physics-inspired machine learning model (QM-ML) for predicting logP and compare its performance with an empirical descriptor-based model and a graph neural network (GNN) model. For all the models investigated, feature selection was conducted to refine the high-dimensional descriptor space and identify critical determinants of lipophilicity or logP. Quantum mechanical descriptors, in particular frontier orbital energies derived, such as chemical potential, hardness, electrophilicity, electronegativity, and dipole moment, transfer energy, and quantum mechanical polar surface area (QMPSA), were extracted to construct QM-ML mod. The results indicate that the hybrid model achieved the highest predictive accuracy, yielding an R2 of 0.9248, a mean absolute error (MAE) of 0.3511, and a root mean square error (RMSE) of 0.4981. The ExtraTrees regressor, utilizing empirical classical Mordred descriptors, demonstrated comparable performance with an R2 of 0.9141. While models relying exclusively on frontier orbital energies showed limited accuracy (R2 = 0.4349), the inclusion of explicit solvation energy (ΔG) and QMPSA significantly improved the physics-based model performance (R2 = 0.8132). The DeepChem-based GCN model also achieved an R2 of 0.8936. Notably, the explainability analysis showed that certain QM descriptors, such as solvation energy (ΔG) and non-polar QMPSA, contributed significantly to the model performance. The hybrid model outperformed all the physics-based and baseline Mordred models with large-scale descriptor sets and showed the importance of certain QM descriptors, such as solvation transfer energy and QM-PSA. These findings highlight the potential of a hybrid data-driven model involving physics-based descriptors to capture complex molecular interactions and underscore the importance of targeted feature engineering in quantitative structure–property relationship modeling.