Biomedical Engineering Reference
In-Depth Information
overall accuracies of 82-86% for the training set and 83-85% for the test set. The
selected models were further tested by using the PubChem database. The models
showed a poor predictivity in the classification of the test set compounds, with only
78 out of 193 compounds correctly classified. In addition, 50 compounds from the
Cambridge database that were predicted as hERG blockers and 10 compounds that
were predicted as nonblockers by two or more models were selected for experi-
mental validation. Of the 50, 18 predicted hERG inhibitors showed more than 50%
displacement of astemizole, while all the predicted nonblockers were found to be
inactive.
The SOM was used by Hidaka et al. [
77
] to classify the compounds using
structural information. First, 37 compounds were divided into three classes
depending on whether the activity (pIC
50
) was below 5, between 5 and 7, or higher
than 7. The analysis of the map obtained reveals that the potent blockers and the
inactive compounds occupy two different areas, while the compounds with the
intermediate activity overlap the two areas. The same method was then applied to
the public available dataset. They divided this database into active (“Hit”) and
inactive compounds. The “Hit” molecules were subdivided into compounds that
cause hERG blockade between 20 and 30%, between 30 and 50%, and more than
50%. Considering only the “Hit” compounds they established a line which divides
the map into two parts: hERG positive and hERG negative areas. All compounds
that block the hERG channel by more than 50% are mapped in the area of the hERG
positive potential, with the exception of one false negative.
4.6 Matched Molecular Pairs
In a recent and interesting study, Papadatos et al. [
78
] applied the matched molecu-
lar pairs technique to three large data sets: hERG (76.266 compounds), solubility
(94.053 compounds), and lipophilicity (180.440 compounds) to find the most
frequent modifications of the molecules. The matched molecular pair analysis of
the hERG database identified 15 frequent transformations. These modifications are
related to only one or two heavy atoms, except for the substitution of a hydrogen
atom with a phenyl ring. In 9 transformations out of 15, the modifications have only
a small or no effect for the hERG affinity of the compounds. According to previous
results, they found that replacing a hydrogen atom with a hydroxy group is
detrimental for the hERG affinity in 45% of the cases, while replacing a hydrogen
atom with a phenyl ring increases the hERG inhibition in 65% of the cases. The use
of context descriptors, such as reduced graphs, Murcko frameworks and Daylight
fingerprints, as well as more local descriptors such as localized RG nodes and atom
environments, highlights significant trends that are not evident when only the
matched molecular pairs technique is used. For example, considering the substitu-
tion of a hydrogen atom with a methoxy group the global distribution indicates that
the possibilities to reduce or to improve the hERG affinity are more or less identical.
However, if one takes into account the reduced graph node the scenario appears to
