Predicting Substrate Properties for P-Glycoprotein

As already outlined above, P-glycoprotein is constitutively expressed in several organs, such as kidney, liver, intestine, and also at the blood-brain barrier. P-gp substrates therefore show poor oral absorption, enhanced renal and biliary excretion, and usually do not enter the brain [28]. This spurred the development of medium- and high-throughput systems addressing the P-gp substrate properties of compounds of interest. These systems mostly rely on transport studies through a monolayer of P-gp expressing Caco-2 [29] or MDCK cells [30]. In parallel, in silico methods have also 

110 are predominantly substrates and those with MolES 49 are nonsubstrates. However, only 30% of the compounds analyzed comply with these two thresholds, all others have values between 49 and 110. 

Cabrera et al. pursued a topological substructural approach for the prediction of P-gp substrates. A linear discriminant model classified 163 compounds with an accuracy of 81% based on standard bond distance, polarizability, and the Casteiger-Marsilli atomic charge [35]. Furthermore, the predictive potential ofthis TOPS-MODE approach was demonstrated for a set of 6-fluoroquinolones not included in the training set. 

On the basis ofthe Cottesman data set and a set of259 compounds compiled from the literature, we explored the performance of several classification methods combined with different descriptor sets. These include simple ADME-type descriptors (log P, number of rotatable bonds, number of H-bond donors, and acceptors), 

When comparing binary QSAR and support vector machines, the latter gave more robust models with total accuracies in the range of 80%. In general, the prediction of nonsubstrates is better than those for substrates [42]. 

---