Aqueous solubility prediction is critical in diverse fields, including drug discovery and chemical engineering, where understanding molecular interactions in solvents can significantly impact formulation and application strategies. This study presents a comprehensive machine learning-based framework for predicting the aqueous solubility of molecules using multiple machine learning models. These include regression- and ensemble-based approaches. Four distinct featurization techniques: Custom descriptor-based defaturization, RDKitDescriptors defeaturizer, extended connectivity fingerprints (ECFP), and MACCS keys were evaluated to transform molecular SMILES strings into numerical features. Three curated datasets (AQSol, ESOL, and PHYS) were utilized for training and evaluation. Along with these datasets, an independent test dataset containing 62 anticancer compounds was also used exclusively for testing purpose. The results showed that molecular descriptors that included physicochemical, structural, and topological properties, outperformed all other methods across multiple metrics ( \( R^2 \) , MSE and RMSE). This study highlights the importance of feature richness and balanced dimensionality in solubility prediction. It also lays a foundation for further exploration of graph neural networks to enhance predictive performance.