Reimagining QSAR Modelling with Quantum Chemistry: A CYP1B1 Inhibitor Case Study

CYP1B1 (Cytochrome P450 1B1) is a key enzyme involved in the metabolic activation of carcinogens and is gaining importance as a therapeutic target, especially in hormone-dependent cancers like breast cancer. In the present work, we have developed QSAR (Quantitative Structure–Activity Relationship) models for 63 previously reported compounds with known inhibitory activity (pIC₅₀) against CYP1B1. Rather than relying on conventional 2D/3D descriptors, we focused on extracting quantum chemical and thermodynamic descriptors using xTB, a semi-empirical quantum tool. Parameters like HOMO-LUMO energy gap, dipole moment, zero-point energy, entropy, and enthalpy were computed from optimized geometries. Using recursive feature elimination (RFE), the top 8 descriptors were shortlisted and used for both regression and classification modelling. Several machine learning techniques were applied, including Multiple Linear Regression (MLR), Support Vector Regression (SVR), Random Forest, and K-Nearest Neighbours for regression, and Support Vector Classification (SVC), ensemble voting, stacking, and XGBoost for classification. Among the models tested, classification models gave better performance compared to regression. The stacking classifier achieved an accuracy of 92.3% with an AUC of 0.86, while the XGBoost model showed comparable accuracy and a slightly higher AUC of 0.94. These findings show that quantum and thermodynamic descriptors, even without conventional structural fingerprints, can provide meaningful insights for activity prediction. This study provides a foundation for further work, where we plan to use docked conformers to capture interaction-based features for improved biological relevance.