Investigation of Effective Molecular Dynamics-derived Properties on Drug Solubility via Machine Learning

Solubility is an essential factor in drug discovery and development, as it plays a vital role in a medication's bioavailability and remedial efficacy. Understanding solubility at the initial stages of drug discovery is essential for reducing resource consumption and improving the probability of clinical success via prioritizing compounds with optimal solubility. Molecular dynamics (MD) simulation is a powerful tool for modeling a range of physicochemical properties, particularly solubility. MD simulations provide a detailed view of molecular interactions and dynamics, thus revealing insights into the factors influencing solubility. This research aims to statistically examine the importance of ten MD-derived properties along with LogP, one of the most influential experimental properties, on drug aqueous solubility through Machine Learning (ML) techniques. For this purpose, a dataset of 199 drugs from diverse classes were compiled from literatures and studied on MD simulation, and relevant properties extracted and selected as features. Additionally, corresponding octanol-water partition coefficients (Log P) from former studies were also imported and considered in this research. By explicit analysis, properties with the highest influence on solubility were identified. They used as input features for four machine learning ensemble algorithms, namely, Random Forest, Extra Trees, XGBoostRF, and Gradient Boosting. The results show that seven properties (Log P, SASA, Columbic_t, DGSolv, RMSD, AVG shell, and LJ) are highly effective in predicting solubility. For the test set, the best estimator algorithm, Gradient Boosting, a predictive R² = 0.87 and RMSE = 0.537 were achieved. This research underscores the potential of integrating MD simulations with machine learning methodologies to improve the accuracy and efficiency of aqueous solubility predictions in drug development.