Background: Assessment of drug cardiac safety is critical in the development of new compounds and is commonly addressed by evaluating the half-maximal blocking concentration of the potassium human ether-à-go-go related gene (hERG) channels. However, recent works have evidenced that the modelling of drug-binding dynamics to hERG can help to improve early cardiac safety assessment. Our goal is to develop a methodology to automatically generate Markovian models of the drug-hERG channel interactions.
Methods: The training and the test sets consisted of 20800 and 5200 virtual drugs, respectively, distributed into 104 groups with different affinities and kinetics to the conformational states of the channel. In our system, drugs may bind to any state (individually or simultaneously), with different degrees of preference for a conformational state and the change of the conformational state of the drug bound channels may be restricted or allowed. To model such a wide range of possibilities, 12 Markovian chains are considered. Our approach uses the response of the drugs to our three previously developed voltage clamp protocols, which enhance the differences in the probabilities of occupying a certain conformational state of the channel (open, closed and inactivated). The computing tool is comprised of a classifier and a parameter optimizer and uses linear interpolation, support vector machines and a simplex method for function minimization.
Results: We propose a novel methodology that automatically generates dynamic drug models using Markov model formulations and that elucidates the states where the drug binds and unbinds and the preferential binding state using data obtained from simple voltage clamp protocols that captures the preferential state-dependent binding properties, the relative affinities, trapping and non-trapping dynamics and the onset of IKr block. Overall, the tool correctly predicted the class of 92.04% of the drugs and the model provided by the tool accurately fitted the response of the target compound, the mean accuracy being 97.53%. Moreover, generation of the dynamic model of an IKr blocker from its response to our voltage clamp protocols usually takes less than an hour on a common desktop computer.
Conclusion: Our methodology could be very useful to model and simulate dynamic drug-hERG channel interactions. It would contribute to the improvement of the preclinical assessment of the proarrhythmic risk of drugs that inhibit IKr and the efficacy of antiarrhythmic IKr blockers.
Keywords: Drug modeling; I(Kr) blocker; In-silico model; Ion channels; Machine learning; hERG blocker.
Copyright © 2022. Published by Elsevier B.V.