Machine Learning-Based Soft Voting Ensemble Model for the Prediction of Oral Drug-Likeness of Chemical Structures

Drug discovery is a costly process with low success rates. Early prioritization of compounds based on drug-likeness estimation can help reduce attrition rates. Here we present HADES (Holistic AI-based Drug-likeness Estimation Score), a soft voting ensemble of five machine learning models that integrates ADMET and physicochemical properties to provide a unified prediction of drug-likeness. Across diverse benchmarks, HADES outperformed other models, more reliably distinguishing nondrugs, clinical-phase candidates, and approved drugs. It assigned low scores to orally toxic, chemically implausible ("odd"), and small-ring-containing structures, highlighting its discriminatory power between drug-like and nondrug molecules. In hit/lead optimization studies, HADES showed systematic score increases following molecular optimization, supporting its use by medicinal chemists to plan molecular optimization campaigns. Furthermore, in virtual screening tasks, HADES delivered high drug-likeness enrichment, enabling efficient prioritization of known active compounds within virtual libraries. Overall, our results indicate that HADES is a practical tool that can be used for compound prioritization and decision-making in early drug discovery.