Prediction of hERG Channel Inhibition Using In Silico Techniques - Ion Channels and Their Inhibitors - page 213

models were built using the dataset divided into class 1 (low hERG activity), class
2 (high hERG activity) and class 3 (medium hERG activity). The dataset
compounds were divided into the training set by random selection (80:20) or by
diverse subset selection (80:20). The architecture of the CPG-NNs was designed
with 3 output layers representing the classes 1-3. The CPG-NN models were
trained with 16 different sets of descriptors. The CPG-NN model obtained using
the set of 11 hERG relevant descriptors and the reference set of 20 diverse inhibitors
showed the best performance, reaching a total accuracy of 0.73-0.74 for the test set
and 0.92-0.93 for the training set. The best Binary QSAR and CPG-NN models
were validated using 1,806 compounds published in the PubChem compound
library and 58 compounds collected from the literature. With the threshold value
of 1

M, the Binary QSAR model achieved the total accuracy of 0.93. The CPG-NN
model correctly classified 68% of the compounds in class 1, 100% of the
compounds in class2 and 75% of class 3.
Jia et al. [ 70 ] designed a classification model using an SVM and the atom type as
molecular descriptors. The model yielded an overall accuracy of 99.59% for the
977 compounds of the training set and of 94% for the 66 compounds of the test set.
The use of the atomic molecular descriptors makes the classification model easy to
interpret. The most important atom-type descriptors were the N16 (nitrogen atom in
an aliphatic ring), C17 (unsubstituted carbon atom next to the N16 in a ring), H4
(acidic hydrogen) and M12 (number of aromatic rings). The nitrogen of the N16
descriptor was usually protonated, and it was found in hERG blockers 308 times out
of 322 times of the total occurrence of the descriptor. The C17 descriptor was found
1,635 times of which 1,318 times it was associated with hERG inhibitors. The H4
descriptor was found in 60 structures of which only 5 were hERG positives, in
agreement with the fact that negatively charged compounds are normally
nonblockers. The M12 descriptor occurred 2,581 times. Only 94 out of 535
molecules containing 3 aromatic rings were hERG inactive, while 14 out 19
compounds containing one aromatic ring were nonblockers.
New descriptors generated from the Shape Signature method were used by
Checkmarev et al. [ 66 ] in combination with k nearest neighbors ( k -NN), SVMs,
and Kohonen self-organizing maps (SOM) to create classification models. The
models were built based on a dataset of 83 compounds divided into strong binders
(IC 50 <

m

M). Two different sets of molecular
descriptors were calculated, one based only on molecular shape and the other one
based on molecular shape and polarity. The SVM models showed a better perfor-
mance than the k -NN with an overall accuracy of 69-74% and of 66-67%,
respectively.

1 m M) and weak binders (IC 50 > 10 m

4.5.1 Decision Trees

The decision tree approach was chosen by several groups to classify hERG blockers
and nonblockers. The ClogP, MR and p K a were used by Buyck et al. [ 71 ].