Computing Techniques for Modeling the Retention of Polar Pharmaceutical Active Compounds by Membranes

Abstract

The retention of polar pharmaceutical active compounds (PPhACs) by nanofiltration (NF) and reverse osmosis (RO) membranes is of paramount importance in membrane separation processes. The retention of 21 PPhACs was correlated using computational intelligence techniques (feedforward neural networks multi-layer perceptron, feedforward neural networks radial basis function, and support vector machines). A database of 541 retention values has been collected from the literature. The results showed a high training and predictive capacity of the FNN-MLP model for the retention of PPhACs by NF/RO with a very high correlation coefficient (R = 0.9714) and a very low root mean squared error (RMSE = 3.9139%) for the all phase, which the FNN-MLP model is better than that obtained using FNN-RBF model (R = 0.7807 and RMSE = 10.2785%) and SVM model (R = 0.9677 and RMSE = 4.1598). The sensitivity analysis was computed and emphasized that the retention of PPhACs is governed by three interactions arranged in descending order: the polarity interactions (hydrophobicity/hydrophilicity) (Relative Importance RI = 75.68%), electrostatic repulsion (RI = 58.88%), and steric hindrance (RI = 54.73% and RI = 32.25%). This research suggests that the PPhACs retention on the NF/RO strongly depend much more on the topological polar surface area.