Biomedical Engineering Reference
In-Depth Information
In total 882 compounds were used by Cianchetta et al. [
40
] to develop a GRIND-
based model. Four probes representing hydrophobic interactions (DRY), hydrogen
bond acceptor (sp
2
carbonyl oxygen), hydrogen bond donor (neutral flat amide
NH), and the molecular shape (TIP) were used to calculate the GRIND descriptors.
The correlation between the GRIND descriptors and the pIC
50
values of the hERG
blockers was analyzed through multivariate techniques such as principal compo-
nent analysis (PCA) and partial least squares (PLS). The dataset was subdivided
into two subsets characterized by the presence or absence of a basic nitrogen atom.
The 338 hERG blockers, which form the subset of compounds without a charged
nitrogen, were subdivided into a training set of 322 compounds and a test set of 16
molecules. The model obtained resulted in four latent variables and showed an
r
2
of
0.76 and a
q
2
of 0.72. The 544 compounds that constitute the charged nitrogen
database were split into a training set of 518 molecules and in a test set of 26
compounds. The model obtained from the PLS regression analysis had three latent
variables yielding an
r
2
of 0.77 and a
q
2
of 0.74. The descriptors involved in the two
models were practically identical. The authors suggest that this might indicate that
the charged and noncharged compounds share the same binding mode. The two
models differ in terms of distance between the edge of the molecule and the space
between a hydrophobic MIF and a hydrogen bond donor group. In the first case,
the distance for the optimum space between the two fields generated is 25
˚
for the
noncharged molecules and 29
˚
for the charged compounds. In the second case,
the DRY/hydrogen bond distance for the noncharged compounds is 14
˚
, while for
the charged molecules it is 21
˚
. Statistical analysis shows that in both models a
hydrogen bond moiety close to the edge of the molecules plays an important role.
Johnson et al. [
35
] analyzed 1,075 compounds through a combination of physi-
cochemical and pharmacophoric descriptors. Least median squares (LMS) regres-
sion was used to analyze the 925 compounds of the training set and the resulting
model shows an
r
2
of 0.65 and a
q
2
of 0.66. In a further test using 1,679 compounds,
the model showed an
r
2
of 0.54. These compounds were then clustered based on
Daylight Fingerprint Tanimoto using the average linkage method and a similarity
cutoff of 0.7. In the largest cluster, the model achieved an
r
2
of 0.32. However,
when only the compounds in the training set with a Tanimoto similarity greater than
0.65 were considered, the
r
2
for the validation set increased to 0.72.
With the aim to develop a model able to discriminate between hERG inhibitors
and noninhibitors, Li et al. [
41
] combined pharmacophore-based GRIND
descriptors and support vector machine (SVM) techniques. Four GRIND probes
representing hydrophobic interactions (DRY), hydrogen bond acceptors (sp
2
car-
bonyl oxygen), hydrogen bond donors (NH neutral flat amide), and molecular shape
descriptors (TIP) were used. From a library of 495 compounds, only the 192
molecules with IC
50
values lower than 40
M were considered. The model
generated from the PLS analysis consists of three latent variables and shows a
rather poor
r
2
of 0.34 and a
q
2
of 0.07. The analysis of the predicted versus the
experimental pIC
50
shows that the IC
50
of sertindole derivatives is overestimated. A
better model with
r
2
m
0.57 and
q
2
0.41 was obtained considering only the
hERG blockers with an IC
50
value lower than 32 nM and by removing the sertindole
¼
¼
