Prediction of drug transfer into human milk from theoretically derived descriptors

Abstract

The goal of this study was to develop a genetic neural network (GNN) model to predict the degree of drug transfer into breast milk, depending on the molecular structure descriptors, and to compare it with the current model. A supervised network with back-propagation learning rule and multilayer perceptron (MLP) architecture was used to correlate activity with descriptors that were preselected by a genetic algorithm. The set of 60 drug compounds and their experimentally derived MIP values used in this study were gathered from Literature. A total of 61 calculated structural features including constitutional, topological, chemical, geometrical and quantum chemical descriptors were generated for each of the 60 compounds. The MIP Values were used as the ANNs output and calculated molecular descriptors as the inputs. The best GNN model with 26 input descriptors is presented, and the chemical significance of the chosen descriptors is discussed. Strong correlation of predicted versus experimentally derived M/P values (R-2>0.96) for the best ANN model (26-5-5-1) confirms that there is a link between structure and MIP values. The strength of the link is measured by the quality of the external prediction set. With the RMS error of 0.425 and a good visual plot, the external prediction set ensures the quality of the model. Unlike previously reported models, the GNN model described here does not require experimental parameters and could potentially provide useful prediction of M/P ratio of new potential drugs and reduce the need for actual compound synthesis and experimental M/P ratio determination.