Classification models for hERG inhibitors by counter-propagation neural networks

Counter-propagation neural networks were used to develop computational models for classification and prediction of human ether-a-go-go-related-gene (hERG) potassium channel blockers. The data set used includes 285 compounds taken from literature sources and two sets of 2D molecular descriptors, one is based on 32 P_VSA descriptors derived from MOE and the other comprises 11 descriptors retrieved by a feature selection method. The counter-propagation neural networks with a 3-dimensional output layer combined with a set of 11 hERG relevant descriptors showed best performance, especially in classifying compounds in the middle-activity class ( hERG IC50 = 1-10 mu M). The total accuracy values obtained for training and test sets are 0.93-0.95 and 0.83-0.85, respectively. In each activity class (low, medium, high), 'Goodness of Hit lists' GH scores archived range from 0.89 to 0.97 for the training set and from 0.74 to 0.87 for the test set. This model thus provides possible strategies for improving the performance of predicting and classifying compounds having hERG IC50 in the range of 1-10 mu M.