Exploring QSTR and toxicophore of hERG K+ channel blockers using GFA and HypoGen techniques

Predictive quantitative structure–toxicity and toxicophore models were developed for a diverse series of hERG K+ channel blockers, acting as anti-arrhythmic agents using QSAR+ module in Cerius2 and HypoGen module in Catalyst software, respectively. The 2D-QSTR analysis has been performed on a dataset of 68 molecules carefully selected from literature for which IC50 values measured on hERG K+ channels expressed in mammalian cells lines using the voltage patch clamp assay technique were reported. Their biological data, expressed as IC50, spanned from 7.0 nM to 1.4 mM, with 7 orders difference. Several types of descriptors including electrotopological, thermodynamic, ADMET, graph theoretical (topological and information content) were used to derive a quantitative relationship between the channel blockers and its physico-chemical properties. Statistically significant QSTR model was obtained using genetic function approximation methodology, having seven descriptors, with a correlation coefficient (r2) of 0.837, cross-validated correlation coefficient (q2) of 0.776 and predictive correlation coefficient (rpred2) of 0.701, indicating the robustness of the model. Toxicophore model generated using HypoGen module in Catalyst, on these datasets, showed three important features for hERG K+ channel blockers, (i) hydrophobic group (HP), (ii) ring aromatic group (RA) and (iii) hydrogen bond acceptor lipid group (HBAl). The most predictive hypothesis (Hypo 1), consisting of these three features had a best correlation coefficient of 0.820, a low rms deviation of 1.740, and a high cost difference of 113.50, which represents a true correlation and a good predictivity. The hypothesis, Hypo 1 was validated by a test set consisting of 12 molecules and by a cross-validation of 95% confidence level. Accordingly, our 2D-QSTR and toxicophore model has strong predictivity to identify structurally diverse hERG K+ channel blockers with desired biological activity. These models provide a useful framework for understanding binding, and gave structural insight into the specific protein–ligand interactions responsible for affinity, and how one might modify any given structure to mitigate binding.