Implementing and tuning machine learning-based models for description of solubility variations of nanomedicine in supercritical solvent for development of green processing

Determination of drug solubility in supercritical solvents such as CO2 has been of great importance for preparation of nanomedicines. This study implements and tunes several machine learning models to describe the solubility of medicine and density of solvent at various pressure and temperature. The dataset used in this study consisted of the input variables, temperature, and pressure. The methods of AdaBoost algorithm to boost the performance of base regression models for predicting the mole fractions of rivaroxaban and the density of SC-CO2 were developed. The base models used include Theil-Sen Regression (TSR), Gaussian Process Regression (GPR), Automatic Relevance Determination Regression (ARD), and Linear Regression (LR). We employ the Hunter-Prey Optimization technique to tune the hyper-parameters of these models. The results indicated that the boosted models outperform their base counterparts. For the mole fraction predictions, AdaBoost with ARD achieves an R2 value of 0.95986, while AdaBoost with GPR obtains an R2 score of 0.99817. For the SC-CO2 density predictions, AdaBoost with GPR achieves an impressive R2 of 0.99906. Accordingly, the AdaBoost with GPR is the best model for both outputs. These results demonstrate AdaBoost strength and Hunter-Prey algorithm in enhancing the predictive accuracy of regression models for these chemical properties.