Selective Fusion of Heterogeneous Classifiers for Predicting Substrates of Membrane Transporters

Membrane
transporters play a crucial role in determining fate of
administered drugs in a biological system. Early identification of
plausible transporters for a drug molecule can provide insights into
its therapeutic, pharmacokinetic, and toxicological profiles. In the
present study, predictive models for classifying small molecules into
substrates and nonsubstrates of various pharmaceutically important
membrane transporters were developed using quantitative structure–activity
relationship (QSAR) and proteochemometric (PCM) approaches. For this
purpose, 4575 substrate interactions for these transporters were collected
from the Metabolism and Transport Database (Metrabase) and the literature.
The transporters selected for this study include (i) six efflux transporters,
viz., breast cancer resistance protein (BCRP/ABCG2), P-glycoprotein
(P-gp/MDR1), and multidrug resistance proteins (MRP1, MRP2, MRP3,
and MRP4), and (ii) seven influx transporters, viz., organic cation
transporter (OCT1/SO22A1), peptide transporter (PEPT1/SO15A1), apical
sodium-bile acid transporter (ASBT/NTCP2), and organic anion transporting
peptides (OATP1A2/SO1A2, OATP1B/SO1B1, OATP1B3/SO1B3, and OATP2B1/SO2B1).
Various types of descriptors and machine learning methods (classifiers)
were evaluated for the development of robust predictive models. Additionally,
ensemble models were developed by bagging of homogeneous classifiers
and selective fusion of heterogeneous classifiers. It was observed
that the latter approach improves the accuracy of substrate/nonsubstrate
prediction for transporters (average correct classification rate of
more than 0.80 for external validation). Moreover, structural fragments
important in determining the substrate specificity across the various
transporters were identified. To demonstrate these fragments on the
query molecule, contour maps were generated. The prediction efficacy
of the developed models was illustrated by a good correlation between
the reported logBB value of a molecule and its predicted substrate
propensity for blood–brain barrier transporters. Conclusively,
this comprehensive modeling analysis can be efficiently employed for
the prediction of membrane transporters of a drug, thereby providing
insights into its pharmacological profile.