Machine learning-assisted computation of the solubility of solid drugs in supercritical carbon dioxide

This research applies three types of machine learning models, i.e., artificial neural network models (radial basis function, multi-layer perceptron, cascade feed-forward, and Elman recurrent neural networks), adaptive neuro-fuzzy inference systems, and least-squares support vector machines, to determine the solubility of twenty-eight solid drugs in supercritical carbon dioxide (scCO₂). These intelligent models estimate the solubility as a function of operating conditions (equilibrium pressure and temperature), properties of solid drugs (critical temperature, critical pressure, acentric factor, molecular weight, and melting point), and scCO₂ density. The relevancy analysis based on Pearson's coefficient suggests that there is a meaningful relationship between the drug solubility in scCO₂ and these input features. Simultaneously applying feature importance analysis with the hyperparameter tuning of machine learning models results in precise modeling for the considered problem. The statistical analyses supported the priority of the multi-layer perceptron (MLP) neural network over other paradigms. This model estimates the testing datasets with the average absolute relative deviation of 16.76%, the mean absolute error of 1.82, the relative absolute error of 12.29%, the root mean squared error of 3.94, and a correlation coefficient of 0.98314. This means that the MLP neural network offered the highest accuracy for predicting the solubility of the considered drugs in the supercritical solvent.