Rethinking the AI Paradigm for Solubility Prediction of Drug‑Like Compounds with Dual‐Perspective Modeling and Experimental Validation

Abstract Aqueous solubility is a crucial property for drug development, as it not only influences the drug delivery process but also determines the bioavailability of drugs. However, solubility prediction remains a formidable challenge, even after decades of research. Most previously‐reported machine learning (ML) models generalize poorly on external sets due to the vast chemical space of drug compounds. In this report, the largest aqueous solubility dataset of drug and drug‐like molecules so far is compiled, based on which reliable models for drug solubility prediction are developed by comparative modelling with assorted regression and classification algorithms. Under current circumstances, even advanced deep learning models are found less accurate than the stacking of multiple statistical ML algorithms due to data limitation. Analysis of applicability domain further verifies the generalization capability of the models for the drug domain, based on which the entries without experimental solubility in the DrugBank database are populated and categorized. Finally, the solubility of ten potential drug molecules is experimentally determined for the first time, again revealing the high reliability of our models. Hence, this work is believed to provides a comprehensive benchmark for future solubility prediction models and a powerful tool to guide new drug discovery.