Accurate estimation of pharmaceutical solubility in green solvent via data-driven structures

Reliable prediction of drug solubility in environmentally benign solvents plays a pivotal role in designing efficient pharmaceutical formulations, improving the biological availability of active compounds, and supporting more sustainable drug-manufacturing practices. In this study, sophisticated data-driven models were developed to characterize and forecast solubility using four key physicochemical and operating parameters: temperature, molecular mass, melting temperature, and system pressure. A comprehensive dataset comprising 355 experimental observations was utilized for model development and validation through a rigorous 5-fold cross-validation framework. Seven machine learning and deep learning approaches, including Decision Tree (DT), AdaBoost, K-Nearest Neighbors (KNN), Random Forest (RF), Ensemble Learning (EL), Multi-Layer Perceptron Artificial Neural Network (MLP-ANN), and Convolutional Neural Network (CNN), were systematically evaluated. Pearson correlation analysis indicated that melting point exhibited the strongest relationship with solubility among the investigated variables (−0.24), whereas molecular weight showed the weakest correlation (−0.08). Furthermore, global SHAP (SHapley Additive exPlanations) analyses identified melting point as the dominant factor affecting solubility predictions, followed by pressure, molecular weight, and temperature. Comparative assessment of the implemented models demonstrated that the CNN architecture delivered the highest predictive capability, achieving a test R 2 of 0.9819 and a remarkably low test MSE of 0.0208, thereby outperforming all competing techniques. In contrast, the MLP-ANN model displayed the weakest generalization performance, yielding the lowest predictive accuracy and the highest error levels among the examined methods. The reliability of the developed frameworks was further assessed through leverage analysis, confirming that the generated predictions remained within a dependable applicability domain without the presence of influential outlying observations. Overall, the proposed interpretable and mathematically robust modeling strategy provides an efficient computational tool for estimating pharmaceutical solubility in green solvents and offers valuable guidance for pharmaceutical formulation design and sustainable process development.