Assessment of drug safety is crucial in the development of new compounds and is typically addressed by evaluating the blockade they cause in the potassium human ether-à-go-go related gene (hERG) channels. It has been recently suggested that the channel state preference for binding and unbinding of a drug also plays a relevant role in drug safety. Our objective is to develop a classifier using machine learning techniques to elucidate the channel state preference of a drug.
We created a set of 2600 virtual IKr blockers with diverse affinities and kinetics that accounted for 13 different preferences to the conformational states of the channel. In order to assess the blocking potency of each in silico drug, three stimulation protocols that enhance the probabilities of the channel to occupy a certain state were applied and the corresponding Hill plots were constructed. For each protocol-drug combination three aspects were studied: IC50, which is the most commonly used parameter for stablishing cardiac safety, the time course of drug-induced hERG inhibition current and an estimation of the time required to reach the steady state block. Nine parameters obtained for each drug were used as input variables to the classifiers in order to differentiate among 13 target classes. A two-step classifier was developed, trained and evaluated. Firstly, we used support vector machines on the IC50 to separate the 13 classes into 3 groups with 4, 5 and 4 classes respectively. Secondly, we used neural networks on each group with all the variables to finally classify the blockers. The three classifiers obtained an overall accuracy on the test group of 90.83, 88.66 and 89.16% for each of the groups respectively. Therefore, our simulations suggest the potential use of our protocols and classifiers to elucidate the channel state preferences of a drug.