Biomedical Engineering Reference
In-Depth Information
and 0.83-0.85 for the test set, respectively. In particular, this model was able to
correctly classify high, moderate, and weak hERG inhibitors with an accuracy of
0.93 for class 1, 0.97 for class 2, and 0.96 for class 3. Using only one output layer,
the CPG-NN was also used to predict hERG affinity. The model based on 11
relevant descriptors showed highest performance with an
r
2
of 0.87 for the training
and the test set.
A new series of fragment/pharmacophore descriptors combined with SVM and
Random Forest (RF) was applied by Catana et al. [
72
] to develop classification
models for a dataset of 561 compounds. An external test set of 1,895 molecules
from the PubChem was used to validate the model. Each molecule was hashed into
different fragments. Subsequently, each fragment was mapped back onto the
dataset compounds, numbered, and a C-fragment descriptor was calculated. The
value of the “comprehensive fragment” descriptor (CF) is calculated by summing
the contribution of each atoms of the fragment. This implies that the values of each
CF descriptor differs for each molecule. The following descriptors were calculated:
E-state (CF_E-state_*), AlogP (CF_AlogP_*), MR (CF_MR_*), positive Gasteiger
partial charges (CF_GC_P_*), negative Gasteiger partial charges (CF_GC_N-*),
the van der Waals surface area positively charged (CF_VSA_P_*), the van der
Waals surface area negatively charged (CF_VSA_N_*), the van der Waals surface
area (CF__VSA_*), the van der Waals surface area with a positive AlogP
(CF_VSA_AlogP_P_*), the van der Waals surface are with a negative AlogP
(CF_VSA_AlogP_N_*). In addition, pharmacophore fingerprints implemented in
MOE were used to generate a new set of descriptors (CP_*) using the approach
described above. The model generated by RF using the CF_Estate_*,
CF_VSA_P_*, CF_VSA_N_*, and CF_GA_N_* and some 2D MOE descriptors
showed an overall accuracy of 0.79 with a precision of 62.2% and 92.3% for true
active and inactive compounds, respectively. For the external test set, the model
correctly predicted 105 of 193 inhibitors and 1,408 of 1,702 inactives. The model
created using the SVM based on the C-pharmacophore descriptors showed a poor
performance for the classification of hERG blockers (78 out of 193 compounds
were correctly predicted). Conversely, the good result achieved in the prediction of
hERG inactives indicates that this model might be used to select nonblockers.
In a recent study, in silico Binary QSAR and CPG-NN were used by Thai et al.
[
63
] to classify hERG blockers and nonblockers. The models were built using a
dataset of 243 compounds with SIBAR descriptors calculated on the basis of four
reference sets: 24 diverse drugs obtained from Sk
old et al. [
75
], 20 hERG blockers,
20 hERG nonblockers, and 20 compounds divided in 10 blockers and 10
nonblockers. The SIBAR descriptors were calculated from 11 selected descriptors
from a total number of 184 2D descriptors, 86 VolSurf descriptors, 50 3D “induc-
tive” QSAR descriptors (related to atomic electronegativity, covalent radii and
intramolecular distances) and 32 P_VSA descriptors. The Binary QSAR models
were generated using 16 SIBAR descriptors and threshold values of 1
€
m
M and
10
M. The best classification model was obtained with the 11 hERG relevant
descriptors using the 20 most diverse hERG blockers as reference set (total accu-
racy of 0.85-0.88 for the training set and 0.73-0.92 for the test set). The CPG-NN
m
