Classification models for HERG potassium channel inhibitors based on the support vector machine approach

Abstract

We calculated 1559 molecular descriptors including constitutional, charge distribution, topological, geometrical, and physicochemical descriptors to characterize the molecular structure of human ether-a-gò-gò related genes (HERG) potassium channel inhibitors. A hybrid filter/wrapper approach combing the Fischer Score (F-Score) and Monte Carlo simulated annealing was used to select molecular descriptors relevant to the discrimination of HERG potassium channel inhibitors. Three classification models with threshold values of IC50 =1.0, 10.0 μmol· L-1, respectively, were built using the support vector machine (SVM) approach. Models developed from 367 training set molecules were validated through 5-fold cross-validation (CV) and the average prediction accuracies were 84.8%-96.6%,80.7%-97.7%, and 87.1%-97.2% for the positive, negative, and overall samples, respectively, which showed better performance than models previously reported in literature. Overall prediction accuracies for the three models using an external test set of 97 molecules were between 67.0% and 90.1%, which were close to or better than the results reported in literature. © Editorial office of Acta Physico-Chimica Sinica.