using MODELLER [97]. Alanine-scanning mutagenesis indicated
that the residues (G648, Y652, V659, and F656) on the S6 trans-
membrane domain and the residues (T623 and V625) on pore
helix had important effects on the interaction with MK-499
(a methanesulfonanilide antiarrhythmic drug). The docking analy-
sis demonstrated that there was aπ-stacking interaction between
Y652, F656, and MK-499. V625 and G648 altered the size or the
shape of the binding pocket and affected the binding with MK-499.
Antihistamine terfenadine and cisapride interacted with Y652 and
F656, but had no interaction with V625. Moreover, F656 engaged
inπ-πstacking interactions with the aromatic groups of most of the
blockers, and Y652 producedπ-cation interactions with the tertiary
nitrogen in ligands [99].
Recently, the crystal structure of the hERG channel without
ligands has been solved at 3.8 A ̊[100]. In this structure, the hERG
is open, while the voltage sensors are in a depolarized conforma-
tion. The central cavity is surrounded by a unique environment,
which may contribute to unusual properties of hERG blocking by
many drugs. The subtle structural rearrangement of selectivity filter
may correlate with hERG’s fast inactivation. It is believed that the
crystal structure of hERG is valuable for designing drugs without
undesirable hERG-related cardiotoxicity.4.2 Ligand-Based
Models for hERG
Ligand-based approaches have been widely used to explore the
structure-toxicity relationship of hERG blockers. The first hERG
pharmacophore model based on 11 antipsychotic drugs and
15 compounds from the literature contained four hydrophobic
features and one positive ionizable feature, producing the R^2
value of 0.9 [101]. Cavalli et al. developed a CoMFA model
based on 31 QT-prolonging drugs, showingR^2 ¼0.952 and
Q^2 ¼ 0.767 [102]. Inanobe et al. used HipHop algorithm to
generate a 3D-QASR model, which contained three hydrophobic
features and one positively ionizable feature [103]. The distance
between hydrophobic features and the positively ionizable feature
ranges from 5.5 to 8.9 A ̊, and the ionizable feature is 1.2 A ̊ apart
from the hydrophobic plane. Most of hERG blockers have a basic
nitrogen center, which is protonated under physiological condi-
tions, and other compounds lacking nitrogen center are referred to
as “neutral” or “uncharged” hERG blockers. Aronov proposed two
pharmacophore models based on 194 uncharged hERG blockers
[104]. These two five-point pharmacophore models contain three
hydrophobic or aromatic features and two hydrogen bond accep-
tors, but the location of the second hydrogen bond acceptor is
different. The author combined them into a six-point pharmaco-
phore model and concluded that ClogP<1, decreasing lipophili-
city, and introducing additional components could reduce hERG
binding for neutral compounds. Moreover, Springer et al. indicated
that increasing polarity, decreasing positive charge, and someMachine Learning-Based Modeling of Drug Toxicity 257