Prediction of busulfan solubility in supercritical CO₂ using tree-based and neural network-based methods

Drug solubility is a critical parameter in the pharmaceutical industry for developing efficient processes for production of nanomedicine at industrial scale. Several attempts have been made in recent years to investigate and obtain this parameter using various data mining methods, including neural networks. In this study, to reduce the error rate in predicting solubility, three methods including Multi-layer Perceptron (MLP), decision tree, and random forest have been applied to 32 rows of experimental data collected from literature for solubility of a model drug in supercritical CO₂. Afterwards, the results of these models are examined and compared with measured data to calibrate and validate the developed models. Finally, the mean squared error improved to 1.77 e −5 in Random Forest Model. MLP and decision tree models mean squared errors are equal to 6.72 e −5 and 3.28 e −5, respectively which is a good result, especially when we can guarantee that the model did not have more problems in predicting the drug solubility and can be used as reliable methods in the pharmaceutical area.