Integrating molecular descriptors and fingerprints via meta-learning for aqueous solubility prediction

Solubility strongly affects a compound’s bioavailability, formulation, and pharmacokinetics, so predicting aqueous solubility is a key challenge in drug discovery and chemical design. We present a machine learning system that uses a stacking-based meta-model and several molecular representations to predict molecular aqueous solubility. The model encodes chemical features, including PubChem fingerprints, MACCS fingerprints, Mol2Vec embeddings, and physicochemical molecular descriptors, to capture a wide range of structural and chemical properties. Using a five-fold cross-validation strategy, seventeen regression models covering linear, kernel-based, instance-based, tree-based, and ensemble learning approaches were methodically assessed. To enhance model performance, hyperparameter optimization was carried out utilizing the Optuna framework. The top-performing models CatBoost, ExtraTrees, Histogram Gradient Boosting, and XGBoost—were chosen as base learners based on benchmarking results, and they were then combined using a stacking regression architecture to provide the final solubility predictions. With high R² values and minimal prediction errors, the suggested model showed great predictive ability across multiple independent external test sets. The framework’s resilience and capacity for generalization were further validated by comparative analysis employing various data splitting techniques. These findings demonstrate how well various molecular encodings can be used with ensemble learning to improve solubility prediction, offering a useful computational tool for molecular property prediction and drug discovery.