Correlation ranking procedure for factor selection in PC-ANN modeling and application to ADMETox evaluation

Abstract

A correlation ranking procedure is proposed for selection of factors in principal component-artificial neural network (PC-ANN). The model was applied in the ADMETox evaluation to predict the carcinogenesis activity of 60 organic solvents and the blood–brain barrier (BBB) partitioning of 115 diverse organic molecules. A total of 150 molecular descriptors, including quantum chemical, constitutional, topological and chemical descriptors were calculated. The resulted descriptors were subjected to principal component analysis (PCA), and a three-layered feed forward artificial neural network (ANN) model was employed to model the nonlinear relationship between the extracted principal components (PCs) and the activities. A correlation ranking procedure is proposed here to select the most relevant set of PCs. First, the nonlinear relationship between each one of the PCs was modeled by separate neural networks and the correlation ability of each PC with the activity data was determined. Then, the PCs were entered to the ANN model based on their decreasing correlation ability. The results supported that the proposed model could predict the carcinogenesis activity and logBBB of the organic compounds with percent relative error lower than 4%. Comparison of the results with two other existing factor selection methods named eigenvalue ranking (EV-PC-ANN) and genetic algorithm (PC-GA-ANN) revealed that the proposed model gave results near to the PC-GA-ANN method, while less accurate results were obtained by the EV-PC-ANN procedure.