Machine learning model for prediction of drug solubility in supercritical solvent: Modeling and experimental validation

We developed a simulation methodology on the basis of machine learning techniques for simulation of pharmaceutical solubility in a supercritical solvent, i.e., CO2 with the perspective of nanodrug production. The X variables considered in this simulation work included pressure and temperature of the system, whereas the response (Y) was considered to be drug solubility in the solvent. The model drug considered in this work is Fenoprofen in which its solubility was modeled at different temperature and pressure to assess its suitability for supercritical processing in nanomedicine preparation via supercritical green technology. The solubility dataset of this work includes 32 data points with two input parameters (temperature and pressure) and one output (solubility in mole fraction) which was assumed to be Y in the computations. Models based on machine learning, including CNN, DNN, and GRNN, were selected as the basis for modeling and analysis performed in this study. Also, the hyper-parameters of these methods have been fine-tuned with the help of an algorithm called Bat algorithm, which comes from nature inspiration. The optimized models with the R2 criterion all have a score higher than 0.99, which shows the significant impact of the Bat algorithm in improving the accuracy of the employed models. Also, based on the calculated RMSE, DNN, CNN, and GRNN have error rates of 7.57E-05, 7.25E-05, and 3.94E-05, respectively. Indeed, GRNN model was finally selected as the research primary model for the solubility prediction, and the max error was reduced to 8.47E-05 using this method.