Biomedical Engineering Reference
In-Depth Information
Compounds were categorized as hERG inhibitors when the following three
conditions are satisfied: clogP
7.3.
A Neural Network model based on the E-state keys and Barnard 4096-bit
fingerprints, and a Bayesian model based on FCFP_6, AlogP, Molecular Weight,
and the counts of hydrogen bond acceptors and donors descriptors were used by
O'Brien et al. [
68
] to generate a consensus model. The models were built using a
dataset of 58.963 compounds randomly divided (80:20) to obtain a training and a
test set of 46.967 and 11.996 compounds, respectively. The Neural Network model
shows slightly better results than the Bayesian model with 85% vs. 82% of
compounds correctly classified. To improve the ability to correctly classify the
compounds the Neural Network and the Bayesian models were combined. The
“recover +ve” classifies the compound as hERG blockers if one of the models
predicts it to be positive. The “recover
3.7,
110
MR
176 and p
K
a
<
ve” classifies a compound as negative if
one of the models predicts it to be negative. With this classification model the
number of false positives increases. The “consensus model” classifies a compound
as positive or negative if both models agree. The model correctly classifies 91% of
hERG blockers and 87% of hERG nonblockers. With the “consensus model” the
rates of false positive and false negative are reduced compared to the Neural
Network and Bayesian models.
A total of 155 descriptors such as physicochemical, topological, SMARTS
strings and SIMAST descriptors were computed on a dataset of 264 compounds
by Gepp et al. [
76
] to generate two decision trees composed by a maximum of eight
branches. The first descriptor used in the two partitioning models is the
pharmacophoric string PHARM$, which correctly classifies 71% of the compounds
in both models and only 13 compounds were misclassified as false positives. The
two models are identical in the two subsequent layers, which contain the descriptors
HACSUR (ratio of surface of hydrogen-bond acceptor atoms to total surface),
T1E (topological electronic index using the number of nonhydrogen atoms), HY
(number of hydrogen atoms), DIPDENS (dipolar density) and T2E (topological
electronic index using the number of bonds between nonhydrogen atoms). The
differences between the two models start from the fourth layer. The first model
contains seven branches and the last fourth layers contain the descriptors HLSURF
(ratio of surface on halogen atoms to total surface), MDE23 (molecular distance-
edge vector
l
23
), MR, CHBBA (covalent hydrogen-bond basicity), logP, QSUMN
(sum of atomic charges on nitrogen atoms), and MGHBD (minimal geometric
distance between two hydrogen-bond donor atoms). In the second partitioning
model, the SIMAST descriptor (fingerprint similarity compared to astemizole) is
used in many branches instead of the MR and logP, and the QSUM- (sum of
negative ESP charges) replaces the MR in the fifth layer. The first and the second
partitioning models achieve overall accuracies for the training set of 91.7% and
93.2%, and of 76% and 80% for the test set.
Recursive partitioning models were developed by Dubus et al. [
62
]. They used
203 molecules from the Aureus Pharma database, 32 P_VSA descriptors and
23 uncorrelated relevant descriptors selected from 184 2D-descriptors calculated
with the Molecular Operating Environment (MOE) software. Model1 used an
