Prediction of hERG Channel Inhibition Using In Silico Techniques - Ion Channels and Their Inhibitors - page 209

features, which might interact with the aromatic amino acids Phe656 and Tyr652.
The model also suggests that the amino group should be located asymmetrically
between the hydrophobic or aromatic features, to interact optimally with the hERG
channel.
A set of linear solvation energy relationship (LSER) descriptors were used by
Yap et al. [ 74 ] to develop an SVM-based classification using a training set of 271
compounds collected from the ArizonaCERT, Micromedex, Drug Information
Handbook , Meyler's side Effect of Drugs, and from the work of De Ponti and the
American Hospital Formulary Service. The model obtained was validated using
leave-one-out and Y-randomization methods and model was further tested with an
independent validation set of 78 compounds. Furthermore, the prediction accuracy
was compared with those obtained from other classification methods such as k
nearest neighbor (KNN), probabilistic neural network (PNN), and C4.5 decision
tree. The SVM classification model shows a higher performance than the other
three classification methods, reaching a prediction accuracy of 97.4% and 84.6%
for the blockers and nonblockers and an overall accuracy of 91%.
Tobita et al. [ 60 ] also generated an SVM model, which achieved a prediction
accuracy of 90% and 95% with two different test sets. To build the discriminant
models they used 73 compounds from the literature for which 57 2D descriptors
from the MOE software package, and 51 molecular fragment-count descriptors
taken from the MACCS key set were calculated. The SVM implemented in the
WEKA software package was used to develop two different discriminant models
using IC 50 values of 1

M to define blockers and nonblockers. The
accuracy of the two models to correctly classify the inhibitors and the noninhibitors
was evaluated through tenfold cross validation. The model shows an accuracy of
prediction of 70% when it was tested with an external dataset of 827 compounds
using a threshold 1

m M and 40 m

M. In both models, the most accurate classification was achieved
selecting eight descriptors. For the model with a threshold of 40

m

M the descriptors
selected were five 2D descriptors (SlogP, PEOE_VSA6, PEOE_VSA

m

þ

1,
SMR_VSA5, DIAMETER), and three fragment-count descriptors (number of
NH 2 fragments, which is correlated with the possible number of hydrogen bond
sites, ACH 2 CH 2 A and A$A!A$A related to flexibility/hydrophobicity of the mole-
cule and the presence of two rings connected by a bond respectively). The fragment
ACH 2 CH 2 A also suggests that the presence of a long chain might play an important
role for potent hERG blockers. In the models with a threshold of 1

M, the
descriptors indentified as important were three 2D descriptors (VSA_BASE,
PEOE_Vsa0, SMR_VSA0), and five molecular fragment-count descriptors
(OAAAO, ACH 2 AAACH 2 A, Nnot%A%A, ACH 2 AACH 2 A and eight-membered
or larger rings). The descriptors selected by the two models are basically different.
In the model with the IC 50 value of 40

m

M, the 2D descriptors selected are measures
of global properties of the compounds, while in the model with the IC 50 value of
1

m

M they are related to a particular structure of the hERG blockers. This highlights
that nonpotent hERG blockers need to satisfy some general properties such as the
DIAMETER and the SlogP, while for potent blockers the presence of certain
structural fragments is important. The analysis of molecular fragments shows that

m