Exploring the solubility potential of anti-cancer and supportive agents in supercritical CO2 through advanced computational intelligence techniques

Abstract

The accurate prediction of solid drug solubility in supercritical carbon dioxide (SC-CO₂) is critical for optimizing pharmaceutical processes, especially in environmentally sustainable drug formulation and purification. This study develops a machine learning (ML) framework for predicting solubility in SC-CO₂ using 744 experimental data points (520 training, 112 validation, 112 testing). Four features—melting point, molecular weight, pressure, and temperature—were used as model inputs. A comparative assessment was performed between conventional regression methods (Linear, Ridge, Lasso, Elastic Net) and advanced ML algorithms, including Support Vector Machine, K-Nearest Neighbors, Decision Tree, Random Forest, Gradient Boosting, XGBoost, LightGBM, CatBoost, Gaussian Process Regression, Artificial Neural Networks, and Convolutional Neural Networks (CNN). The results show that tree-based ensembles and deep learning approaches significantly outperform linear models. Notably, the CNN model achieved the best test performance with R² = 0.9839 and MSE = 0.0800, followed by CatBoost (R² = 0.9795) and Gaussian Process Regression (R² = 0.9751). Feature importance analysis using SHAP revealed molecular weight as the most influential variable, followed by pressure, temperature, and melting point. Overall, this study highlights the potential of ML in improving solubility prediction and supports its application in early-stage drug development and green pharmaceutical processing.