Advances in Machine Learning Approaches for Predicting Aqueous Solubility in Drug Discovery

Abstract

Nearly 90% of drug candidates and 40% of licensed medications have poor water solubility. Aqueous solubility is a critical factor in drug discovery and development, influencing absorption, bioavailability, and therapeutic effectiveness. Machine learning (ML) approaches, including Random Forest, Support Vector Machines, deep learning models, and hybrid techniques, have demonstrated superior accuracy in predicting solubility compared to traditional quantitative structure-property relationship (QSPR) models. These models utilize molecular descriptors, fingerprints, and advanced computational techniques to enhance predictive performance, enabling efficient screening of large chemical libraries. Challenges such as data quality, model interpretability, and overfitting persist, necessitating the adoption of explainable AI, active learning, and transfer learning to improve robustness and generalizability. The integration of ML-based solubility prediction into drug development pipelines has shown promise in optimizing formulation strategies and reducing late-stage failures. This study aims at providing a detailed review on the prediction of aqueous solubility in drug discovery, using machine learning approaches and also the advances found in them. Future research will focus on expanding high-quality datasets, refining hybrid ML-physics models, and leveraging quantum computing to further advance solubility prediction and accelerate pharmaceutical innovation.