Prediction of the hERG potassium channel inhibition potential with use of artificial neural networks

摘要

New drug development is a complex and time consuming process. The procedure is informally divided into strongly inter-dependent phases beginning from chemical structure synthesis or bioengineering through ADME-Tox properties assessment, clinical trials, up to the market introduction. Recently more and more effort has been invested in the early toxicity assessment of the drugs being developed. Apart from relatively well known and widely researched groups of effects which include hepatotoxicity, immunotoxicity, genotoxicity new toxic effects have become deeply investigated. One of the possible and potentially dangerous cardiotoxic effects is triggered by drugs acquired long QT syndrome (LQTS) which can lead to the fatal ventricular arrhythmia what effected in withdrawal of several drugs from the market. In most drugs known causing the ECG (electrocardiography) interferences the effect results from inhibition of the fast potassium channels (encoded as hERG follow the gene name-human ERG). Therefore early prediction of the hERG channel-drug interaction potential has become a major pharmacological safety concern. The objective of this research was to develop a reliable empirical model for the - triggered by drugs - potassium channel inhibition prediction with use of the previously published and publicly available database. The input data consisted of in vitro research settings, drug chemical structure (molecular fingerprints) and physico-chemical parameters for all substances present in database, Artificial neural networks were chosen as the algorithms for the models development and back-propagation multi-layer perceptions as well as neuro-fuzzy systems of Mamdani MISO (multiple input single output) type were tested. Classifiers were built on the training set containing 447 records, describing 175 various chemicals. Two test procedures were applied for the model performance assessment: standard 10-fold cross validation procedure and validation based on the external test set containing 45 records describing. Various activation functions were tested including hyperbolic tangent, sigma and fsr function. The performance of the best model estimated in 10-fold CV was 76% (78% for positive and 74% for negative output respectively). Neural network model with 3 and 2 cells in hidden layers respectively and sigma activation function properly predicted 89% instances from the external validation dataset. By neural model analysis it was possible to estimate quantitative relationship between cardiotoxicity risk potential of particular drug and its lipophilicity described as the log P value.