Prediction of hERG Channel Inhibition Using In Silico Techniques - Ion Channels and Their Inhibitors - page 204

which is based on 20 descriptors selected through the simulated annealing tech-
nique, shows an r 2 of 0.76. The model, tested with an external validation set of
40 compounds, achieves an r 2 of 0.52. Each of the 20 descriptors was used alone to
build a model and the cross-validation r 2 values were calculated. As in many other
models, the top scoring descriptors, such as AlogP and L-10-L (which describe two
hydrophobic moieties separated by a chain of the atoms) shows that hydrophobic
interactions and -stacking are important for hERG blocking. The highest r 2 value

was obtained with the model generated using only the descriptor F-C 1 sp 2 (fluorines
connected to sp 2 carbons). The importance of this group might depend on its ability
to form strong p -stacking interactions. The Goose-Crippen atom type O in phenol,
enol and carboxyl OH shows a negative correlation with the hERG inhibition. It is
noteworthy that two of the three compounds, which show the largest error contain a
carboxylic moiety, thus the model seems to underestimate its contribution to the
decrease of hERG inhibition.
Ekins et al. [ 51 ] showed that also recursive partitioning models can be applied for
early discovery of potential hERG blockers. Ninety-nine compounds from literature
sources were used to develop two models using the ChemTree software package.
In the first model 99 compounds were used to generate 100 random models using
564 path length molecular descriptors. The model showed a good IC 50 correlation
with an r 2 of 0.90, but the prediction of the 35 compounds of the validation test set
resulted in an r 2 of 0.33. In the second model, 134 compounds were used to generate
694 path length descriptors, which were employed in the generation of 100 random
models. The second model showed a lower r 2 value (0.85).
Leong et al. [ 52 ] used a combination of pharmacophore modeling and SVMs to
predict the potency of hERG blockers. The model obtained from the 26 compounds
of the training set had an r 2 of 0.97, while for the 13 compounds of the validation
test set the q 2 value was 0.94.
Recently, Gavaghan et al. [ 53 ] studied 1,312 compounds through the combina-
tion of D-optimal onion design, PLS, and PCA techniques. The molecular
descriptors calculated with the software packages Selma, DRONE, and VolSurf
were integrated with fragment-based descriptors, which were used to calculate base
level PLS models from each descriptor set. The scores generated from the base level
PLS models were combined and used as descriptors in the upper level hierarchical
PLS models. The best hierarchical PLS model generated from 13 descriptors and a
training set of 436 compounds showed a q 2 of 0.59. The model was tested with an
additional external test set of 7,520 compounds, composed of 4,813 actives and
2,707 inactives. Unfortunately, the model was unable to distinguish between potent,
moderate, and nonactive blockers.
Recently, a dataset of 68 compounds collected from the literature was used by
Garg et al. [ 37 ] to perform a 2D-QSTR analysis based on S_sNH2 (primary
amines), JX (Balaban index), Kappa-3 (third order of Kier shape index),
ADMET_PPB (tendency to bind to plasma protein), Atype_O_57 (atom type
Oin phenol, enol, carboxyl OH), Atype_O_59 (atom type O in Al-O-Al) and
Atype_H_46 (atom type H attached to Csp 0 ) descriptors. The model shows a
good prediction ability achieving an r 2 of 0.84 and a q 2 of 0.78. In a validation

p