Prediction of complexation constants and binding free energies of β-cyclodextrin inclusion complexes using machine learning methods

Accurate prediction of complexation constants for cyclodextrin inclusion complexes remains a challenging task due to experimental limitations and the high computational cost of theoretical approaches. In this study, machine learning methods were applied as computational tools to predict the complexation constants of β-cyclodextrin inclusion complexes based on experimentally derived data. A curated dataset of β-cyclodextrin–guest complexes measured at 273 K and pH 7 was combined with a comprehensive set of classical molecular descriptors and SMILES-derived Iso2vec embeddings. Classical descriptors were calculated using the Materials Studio software and included fragment-based structural counts and surface-related Jurs descriptors, while Iso2vec embeddings provided an additional representation of molecular structure and stereochemistry. Feature selection was performed using Heat Map–Based Feature Ranking, enabling the identification of compact and informative feature subsets. Several regression models were evaluated, including linear models and tree-based ensemble methods. Among them, gradient-boosted decision tree models, particularly LightGBM, demonstrated the best predictive performance. The inclusion of Iso2vec embeddings consistently improved model accuracy across architectures, indicating that these features capture structural information not accessible through conventional descriptors alone. Model interpretability analysis using SHAP values revealed that both classical descriptors and Iso2vec components contribute to the final predictions. The proposed approach offers a practical and interpretable framework for data-driven prediction of cyclodextrin complexation constants and may support early-stage decision-making in cyclodextrin-based pharmaceutical formulation development.