Development and evaluation of an effective solubility prediction model for pharmaceuticals in organic solvents using machine learning based on eXtreme Gradient Boosting.

In this work, we have examined the predictive capability of a machine leaning model based on the XGBoost framework as regards the solubility of active pharmaceutical ingredient-like molecules in organic solvents over a wide range of temperatures. A total of 30 binary mixtures has been investigated. The dataset was divided in two sets, with one set for training, testing and validation including solubility data for four solute compounds (butyl paraben, fenofibrate, risperidone, fenoxycarb) consisting of a total of 224 data points, and the second set used for prediction consisting of the solubility data for butamben, with 50 data points in total. The calculated root mean square errors (RMSLE) for the calculated solubility (train, test, validation) were 0.05, 0.09, 0.13 and 0.15, respectively, while the average RMSLE for the predicted solubility of butamben was 0.41. A total of 10 descriptors were considered in this work, comprising parameters for solute (heat of fusion, melting temperature, heat capacity and Hansen solubility parameter), two descriptors representing the solvent (dielectric constant and boiling temperature) as well as temperature, and drug and solvent names. The temperature-dependence of solubility has been captured accurately by setting a constraint on the XGBoost algorithm. A comparison between the performance of the machine learning model proposed and evaluated in this work, and the semi-predictive Flory-Huggins and the temperature-dependent NRTL-SAC models on the other hand, shows that for all the studied compounds, the machine learning model can deliver significantly improved capability to model as well as predict solubility.