P-glycoprotein (P-gp) is an ATP-dependent membrane transporter that plays a pivotal role in eliminating xenobiotics by active extrusion of xenobiotics from the cell. Multidrug resistance (MDR) is highly associated with the over-expression of P-gp by cells, resulting in increased efflux of chemotherapeutical agents and reduction of intracellular drug accumulation. It is of clinical importance to develop a P-gp inhibition predictive model in the process of drug discovery and development.
An in silico model was derived to predict the inhibition of P-gp using the newly invented pharmacophore ensemble/support vector machine (PhE/SVM) scheme based on the data compiled from the literature. The predictions by the PhE/SVM model were found to be in good agreement with the observed values for those structurally diverse molecules in the training set (n = 31, r2 = 0.89, q2 = 0.86, RMSE = 0.40, s = 0.28), the test set (n = 88, r2 = 0.87, RMSE = 0.39, s = 0.25) and the outlier set (n = 11, r2 = 0.96, RMSE = 0.10, s = 0.05). The generated PhE/SVM model also showed high accuracy when subjected to those validation criteria generally adopted to gauge the predictivity of a theoretical model.
This accurate, fast and robust PhE/SVM model that can take into account the promiscuous nature of P-gp can be applied to predict the P-gp inhibition of structurally diverse compounds that otherwise cannot be done by any other methods in a high-throughput fashion to facilitate drug discovery and development by designing drug candidates with better metabolism profile.
Citation: Leong MK, Chen H-B, Shih Y-H (2012) Prediction of Promiscuous P-Glycoprotein Inhibition Using a Novel Machine Learning Scheme. PLoS ONE 7(3): e33829. https://doi.org/10.1371/journal.pone.0033829
Editor: Arto Urtti, University of Helsinki, Finland
Received: May 19, 2011; Accepted: February 20, 2012; Published: March 16, 2012
Copyright: © 2012 Leong et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: Funding was provided by the National Science Council, Taiwan (URL: http://web1.nsc.gov.tw/), grant number NSC99-2113-M-259-005-MY3. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
P-glycoprotein (P-gp), which belongs to the ATP-binding cassette (ABC) super family of transporters, utilizes the energy that is released during the hydrolysis of ATP to actively translocate a wide range of structurally unrelated compounds across the cell membrane . P-gp, which is encoded by human MDR1 (ABCB1) gene and localized to chromosome 7q21, can be found in a variety of normal human tissues, including liver, kidney, small and large intestines, pancreas, brain, ovary and testes –. It is believed that P-gp-mediated efflux plays an essential role in cellular protection as well as in secretion and/or disposition by extruding xenobiotics from mammalian cells . For instance, it has been found that oral absorption and central nervous system entry of various drugs can be limited by the P-gp expression in gastrointestinal tract (GIT) and brain capillary endothelial cells, respectively . As a result, P-gp exerts profound effects on the absorption, distribution, metabolism, excretion and toxicity (ADME/Tox) of an administrated drug .
In addition to expression in normal tissues, P-gp is also widely expressed in many human cancers, causing multidrug resistance (MDR), in which a given non-drug resistant cell or cell line becomes cross-resistant to other diverse drugs after being treated by a single drug. This will result in the reduction of intracellular drug accumulation by active extrusion of drugs from the cell . For example, the efficacy of a variety of antitumor agents, such as doxorubicin, paclitaxel, etoposide and vincristine, is diminished once the tumor cells overexpress P-gp . Furthermore, there is a healthy body of studies to support the fact that P-gp plays a critical role in drug resistance in infectious diseases , , brain diseases , rheumatoid arthritis  and cancers , resulting in impairing chemotherapeutic treatment. For instance, 17-allylamino-17-demethoxygeldanamycin (17-AAG) is the first-generation inhibitor of molecular chaperone heat shock protein 90 (Hsp90), which has been proposed to be a novel therapeutic target for a variety of cancers  because of its pivotal role in cancer progression and tumor survival . Nevertheless, the efficacy of 17-AAG is limited by its sensitivity to MDR . As such, novel Hsp90 inhibitors that can inhibit P-gp are under clinical development –. Thus, MDR can increase efflux of chemotherapeutical agents, reduce intracellular drug accumulation, and create a supreme hurdle in the effective chemotherapy of many disorders .
Inhibition of P-gp have broad and profound drug metabolism and pharmacokinetics (DM/PK) implications  since it can either reduce the hepatic and renal clearance or increase the bioavailability, resulting in adverse drug–drug interactions . Thus, some MDR modulators may alter not only the concentration of chemotherapeutic agents in cells, but also their plasma concentrations. For example, a clinical study unequivocally demonstrated that the plasma concentration of orally administered digoxin was dramatically reduced in combination with co-administrated rifampin due to the P-gp mediated drug–drug interactions . Therefore, the inhibition of P-pg plays a clinically important role in modern chemotherapy since it is hoped to find specific P-gp modulators that can efficaciously reverse MDR in resistant cell lines, restore sensitivity to chemotherapy, and thus improve treatment results .
In silico approach has been proven to be a feasible and efficient way to drug ADME/Tox assessments . Of various modeling techniques, pharmacophore modeling, which develops a predictive model based on the combination of chemical features to mimic the interactions between ligands and the target protein, is often adopted . In fact, numerous pharmacophore hypotheses have been proposed to predict the P-gp inhibition –. Nevertheless, it is believed that P-gp is a highly flexible protein  as manifested by the fact that it can interact with a broad range of structurally and functionally diverse compounds , . The highly promiscuous nature of P-gp that is a common characteristic of membrane proteins  can be further illustrated by the published crystal structures of the bacterial lipid transporter MsbA  and homology models , . Furthermore, the mouse P-pg, whose sequence shares 87% identity with human P-gp, is also highly flexible as demonstrated by Figure 1, in which the crystal structures , unbounded (PDB code: 3G5U) as well as co-complexed with QZ59-RRR (PDB code: 3G60) and QZ59-SSS (PDB code: 3G61), are superimposed. These proteins exhibit significant structural discrepancies, especially the amino acid residues Tyr303, Phe332, Phe339, Phe724, Leu758, Phe974 and Tyr949. In addition, promiscuity is not only the hallmark of P-gp conformation but also its inhibitors since it has been observed that P-gp can have multiple binding sites, viz. polyspecificity , , suggesting that inhibitors can interact with P-pg using different chemical features.
The superposition of three murine P-gp proteins, whose PDB codes are 3G5U, 3G60 and 3G61 and color-coded by gray, green and maroon, respectively.
Accordingly, no single predictive model will suffice to accurately describe the interactions between this promiscuous protein and those highly diverse inhibitors , otherwise the derived predictive models can only be applied to some specific chemotypes, which, in turn, will produce substantial prediction errors once the test molecules are located outside the domain of those chemotypes. This perplexing situation can be further illustrated by the P-gp substrates, whose binding sites are blocked by most of inhibitors  despite of the fact that substrates and inhibitors can have different binding regions . There is a growing consensus in favor of using pharmacophore ensemble to model the interactions between P-gp and substrates in order to take into consideration its promiscuous nature , , suggesting that it is plausible to accurately model the interactions between P-gp and inhibitors using pharmacophore ensemble.
Nevertheless, the promiscuous nature of P-gp and its inhibitors can be resolved using a novel scheme recently derived by Leong , in which a panel of plausible pharmacophore hypothesis candidates were adopted to construct a pharmacophore ensemble (PhE), which, in turn, was treated as input for regression analysis via support vector machines (SVM) and the PhE/SVM scheme can be illustrated by Figure 3 of Chen et al . Unlike any other analog-base modeling scheme, each pharmacophore member in the PhE symbolizes a single protein conformation or a group of spatially similar protein conformations. As such, the promiscuous nature of target protein can be taken into consideration and, practically importantly, it has been shown that the PhE/SVM model executed better than the consensus prediction of multiple pharmacophore models  Consequently, a number of systems, whose target proteins are highly promiscuous, were also accurately modeled, including the case studies of the liability of human ether-á-go-go related gene (hERG)  as well as CYP2A6–  and CYP2B6–substrate interactions . Additionally, the developed PhE/SVM model revealed a possible new protein conformation that was never reported before in the investigation of CYP2A6–substrate interactions , and it performed better than the pharmacophore ensemble . The aim of this investigation was to develop an accurate, fast and robust in silico model based on the PhE/SVM scheme to predict the binding affinity of P-gp inhibitors. This shall facilitate drug discovery and development by designing drug candidates with better metabolism profile.
Materials and Methods
To construct quality data for this investigation, comprehensive literature search was carried out to retrieve EC50 values of 130 compounds, which were compiled from different source , –, to maximize the structural diversity. In order to warrant a better consistency, the average values were taken in case there were two or more EC50 values in very close range for a given inhibitor. Furthermore, all chemical structures were examined and only those with definite stereochemistry were enrolled. All molecules assembled in this investigation and references to the literature are listed in Table S1 (Supporting Information).
The conformational flexibility of studied molecules was taken into account by creating multiple conformers since three-dimensional conformations of ligands are of critical importance in developing pharmacophore models . As such, all selected molecules were subjected to conformation search to generate the low-lying conformations, which were carried out using the mixed Monte Carlo multiple minimum (MCMM) /low mode  by MacroModel (Schrödinger, Portland, OR). MMFFs  was chosen as force field and the truncated-Newton conjugated gradient method (TNCG) was set as the energy minimization method. Furthermore, the hydration effect and the solvation effect were taken into consideration by using the GB/SA algorithm  and water as solvent with a constant dielectric constant, respectively. The number of selected unique structures was up to 255 with an energy cutoff of 20 Kcal/mol (or 83.7 KJ/mol).
The chemical and biological characteristics of selected samples in the training set play a pivotal role in determining the predictivity of a generated pharmacophore hypothesis, which can be manifested by the fact that different compound selections can produce different pharmacophore models . The critical factor to constructing a perfect training set is to let HypoGen, which was the program employed for automatic pharmacophore generation (vide infra), “learn” new knowledge from the input. For examples, structurally similar compounds with significantly different biological activities or structurally distinct compounds with similar biological activities are expected to serve as perfect entries. Conversely, any redundancy in the predictive models, viz. overfitting or overtraining, can be yielded when structurally similar compounds with similar biological activities are selected as the training set.
Ideally, an ideal training set should consist of at least 16 molecules to warrant its statistical significance, at least 4 orders of magnitude in biological activity, approximately equal compounds in each order of magnitude and novel information concerning structure-activity relationship. More detailed selection criteria have already been discussed elsewhere , .
Thirty-one molecules, which totally consisted of 7142 conformations, were deliberately selected from all collected molecules by visually scrutinizing their chemical structures and activities to constitute the training set for automatic pharmacophore generation and regression and their associated biological activities spanned 7 orders of magnitude. The generated hypotheses were, in turn, validated by those remaining eighty-eight molecules, whose biological activities varied over 5 log units. In addition, those molecules assayed by Labrie et al.  were deliberately designated as the outlier set to assess the extrapolation capacity of the developed model, viz. the level of robustness, since those samples can mimic the real challenges to a predictive model in real situation. Table S1 lists molecules selected for the training set, test set and outlier set and their corresponding pEC50, respectively.
The HypoGen module in Discovery Studio (Accelrys, San Diego, CA) was employed for automatic pharmacophore generation. It produces and ranks the pharmacophore hypotheses, which quantitatively correlate the three-dimensional arrangement of selected chemical features mapped onto those molecules in the training set with the corresponding activities through three phases, namely construction, subtraction and optimization as compared with any other QSAR techniques , , which normally rely on regression to generate predictive models. During the construction phase, HypoGen generates common conformational alignment among those most active molecules in the training set. The less useful pharmacophore hypotheses such as common to most inactive molecules are eliminated from the collection in the subtractive phase. The survived pharmacophore hypotheses are further improved using the stimulated annealing scheme in the optimization phase. The theory and principle of HypoGen have been describe in detail elsewhere .
Hydrogen bond donor (HBD), hydrogen bond acceptor (HBA) and hydrophobic (HP) chemical features, which depict the intermolecular interactions between an H atom on the ligand and a highly electronegative atom such as an O, N or F atom on the protein, between a highly electronegative atom on the ligand and an H atom on the protein and between nonpolar moieties on both ligand and protein, respectively, were chosen for pharmacophore hypothesis development using different feature combination and minimum, maximum and total numbers for each selected chemical feature as well as total features. In addition, the chemical feature weights and tolerances were varied in order to maximize the hypothesis diversity.
Each single predicted pEC50 value by those pharmacophore hypotheses in the PhE was fed as the input of SVM for further regression. In other words, those predicted pEC50 values were treated as descriptors for QSAR model development. As such, the dimensionality of the SVM input space corresponds to the number of pharmacophore models in the ensemble. Furthermore, the regression calculations were carried out by the SVM package LIBSVM (software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm) using the svm-train module and the developed SVM models, in turn, were validated by those compounds in the test set using the svm-predict module. The runtime parameters, namely cost C, the width of the kernel function γ and ε and ν in case of ε-SVR and ν-SVR regression modes, respectively, were automatically scanned using an in-house perl script by the systemic grid search algorithm .
A number of statistical parameters, namely the correlation coefficient (r2) between the predicted and observed values, standard deviation (s), root-mean-square error (RMSE), maximum residual (ΔMax) and mean absolute error (MAE) were used to evaluate the predictivity of a built model. A 10-fold cross-validation scheme, yielding the cross-validation coefficient (q2), was also employed for internal validation.
All generated models were subjected to validations by those criteria, which were initially proposed by Golbraikh et al.  and adopted by Development of Environmental Modules for Evaluation of Toxicity of pesticide Residues in Agriculture (DEMETRA) , shown as follows,where and k are the correlation coefficient and slope of the regression line (predicted vs. observed values) through the origin, respectively, and is the correlation coefficient of the regression line (observed vs. predicted values) through the origin.
Furthermore, the newly proposed modified version of r2 , which is defined as follows,was also adopted to evaluate the quality of a predictive model, which should be large than 0.5 to be an acceptable model.
Of all generated pharmacophore models using various selections of chemical features and runtime parameters, three hypotheses, denoted by Hypo A, Hypo B and Hypo C (listed in File S1), were assembled to construct PhE based on their prediction performances on every single molecule in the training set and test set as listed by Table S1 and their corresponding statistical evaluations as listed by Table 1. These three candidate models in the ensemble consist of a variety of combinations of chemical features, namely one HBD and four HPs in Hypo A; one HBA, one HBD and two HPs in Hypo B and one HBD and three HPs in Hypo C.
In addition to various combinations of chemical features in these three pharmacophore models, their spatial arrangements are also different as exhibited by Figure 2. It can be found that one HBD and two HPs are common features among them and the closest distance between one HBD and one HP and that between two HPs are 6.374 Å and 8.716 Å in Hypo A, respectively, whereas the same measurements vary to 7.081 Å and 10.365 Å in Hypo B as well as 6.506 Å and 8.515 Å in Hypo C, respectively. The discrepancies among these three models can also be rendered by the bond angle centered at one HP and connecting to one HBD and another HP varies from 55.7° in Hypo A to 63.2° and 50.6° in Hypo B and Hypo C, respectively. Figure 3 demonstrates the superposition of these three models, and it can be observed that these three models are different not only in absolute coordinates but also in the relative relationships.
Generated pharmacophore models (A) Hypo A, (B) Hypo B and (C) Hypo C, in which hydrophobic, hydrogen-bond acceptor and hydrogen-bond donor features are represented by light blue blobs, magenta blobs and arrows, and green blobs and arrows, respectively. The interfeature distances and angles among features, depicted in white, are measured in Ångstroms and degrees, respectively.
Superposition of three pharmacophore models Hypo A, Hypo B and Hypo C, denoted in red, blue and green, respectively.
The three pharmacophore models, in general, predicted those molecules in the training set well as asserted by their less significant residuals (Table S1) and their corresponding statistical evaluations, namely parameters RMSE, MAE and s (Table 1). In addition, all of the correlation coefficients, viz. r2 values, are larger than 0.80, suggesting their statistical significance, which can be further confirmed by inspecting the scatter plot of observed vs. predicted pEC50 values as illustrated in Figure 4.
Observed pEC50 vs. the pEC50 predicted by Hypo A, Hypo B, Hypo C and SVM model for those molecules in the training set. The solid line, dashed lines and dotted lines correspond to the SVM regression of the data, 95% confidence interval for the SVM regression and 95% confidence interval for the prediction, respectively.
The maximum residuals in the training set generated by Hypo A and Hypo B were resulted from the prediction of 17 with values of −1.06 and −1.34, respectively, whose residual was only −0.76 by Hypo C. On the other hand, the prediction residuals of 50 were only −0.15 and −0.58 by Hypo A and Hypo B respectively, whereas Hypo C produced the maximum deviation of −1.00. Conversely, 84 was perfectly predicted by Hypo A, Hypo B and Hypo C with only residuals of 0.15, 0.00 and 0.13, respectively. When applied to 89, Hypo A only yielded a residual of −0.15 and Hypo B and Hypo C showed modest errors of 0.44 and 0.33, respectively. Nevertheless, these three models adopted different conformations to bind to P-gp as illustrated in parts A–C of Figure 5, and this discrepancy becomes more pronounced by the superposition of these three conformations as depicted in part D of Figure 5, which clearly illustrates the need to construct a PhE to address the variations in protein conformation.
Pharmacophore models (A) Hypo A, (B) Hypo B and (C) Hypo C fitted to 89 and (E) overlay of these three models, which are color-coded by red, blue and green, respectively. The chemical features are described in Figure 2.
These three hypotheses in the PhE, in general, also executed well for those molecules in the test set as shown in Tables S1 and 1 and Figure 6, which displays the scatter plot of observed vs. predicted pEC50 values for those molecules in the test set. Therefore, it can be affirmed that Hypo A, Hypo B and Hypo C are qualified to constitute PhE based on the their performances in the training set and the test set as well as their statistical evaluations as mentioned above despite the fact that modest performance deteriorations from the training set to the test set can be observed as suggested by all statistical parameters. The r2 value evaluated by Hypo A, for example, was lowered to 0.73 in the test set, viz. a decrease of 0.12 from the training set. Similar observations can also apply to Hypo B and Hypo C. Similar to those observations found in the training set, prediction discrepancies among these three pharmacophore models can also be found in the test set. For instance, Hypo C produced the maximum error from the prediction of 82 with an error of 1.58, whereas Hypo A and Hypo B only yielded residuals of 0.07 and 0.67, respectively. In fact, all of these three models showed various levels of overtraining, albeit marginally, as depicted by their decreases in r2 values and other parameters from the training set to the test set.
Observed pEC50 vs. the pEC50 predicted by Hypo A, Hypo B, Hypo C and SVM model for those molecules in the test set. The solid line, dashed lines and dotted lines correspond to the SVM regression of the data, 95% confidence interval for the SVM regression and 95% confidence interval for the prediction, respectively.
The final PhE/SVM model was generated by the SVM regression of those three pharmacophore hypotheses in the ensemble, yielding the number of the SVM input components (dimensionality) three. The optimal parameters for running SVM, which were selected based on the prediction results of those samples in the training set and cross-validation as listed in Table S1, are summarized in Table 2. It can be observed that the PhE/SVM model executed better than all of those individual hypotheses in the PhE for those molecules in the training set as further demonstrated by the scatter plot of observed vs. the predicted pEC50 values as shown in Figure 4, in which those points obtained from the SVM model are generally closer to the regression line than those obtained from the Hypo A, Hypo B and Hypo C. As a result, the PhE/SVM yielded the largest r2 and the smallest RMSE, ΔMax, MAE and s among those four predictive models (Table 1). In addition, it can also be observed that the PhE/SVM model yielded residuals, which are smaller than the maximal errors produced by those hypotheses in the PhE for most of molecule in the training set and the smallest in some cases, suggesting that the PhE/SVM model is the most accurate model. The predictions of 2 by Hypo A, Hypo B, Hypo C and PhE/SVM, for example, gave rise to residuals of 0.30, 0.48, −0.28 and −0.02, respectively.
When subjected to 10-fold cross-validation, the PhE/SVM model yielded the correlation coefficient q2 of 0.86, which only decreased from the parameter r2 by a value of 0.03, viz. a tiny difference between both correlation coefficients. Thus, it can be asserted that this PhE/SVM model exhibits highly statistical significance between the predicted values and the input data and, more importantly, it is highly possible that this SVM model is a statistically authentic model.
When applied to those molecules in the test set, PhE/SVM only shows negligible performance decreases from the training set as compared with all models in the PhE, which can be depicted by the parameters r2, ΔMax and MAE (Table 1). The MAE value, for instance, only raised from 0.29 in the training set to 0.30 in the test set despite of the fact that the sample size in the latter was ca. 2-fold more than that in the former. In fact, the parameters RMSE and s indicate that PhE/SVM executed better in the test set than in the training set. Thus, it can be assured that the PhE/SVM model is a better predictor than any of pharmacophore models in the ensemble for those molecules in the test set as shown by Figure 6. Most importantly, those negligible differences between both r2 values and between r2 and q2 values as well as the small and consistent RMSE values in both sets manifest the fact that PhE/SVM is a well-trained predictive model since it will otherwise produce at least one substantial difference in case of overtraining.
Eleven molecules, whose inhibition activities of P-gp were investigated by Labrie et al. , were deliberately selected as the outliers to further challenge the extrapolation power of generated models since they are completely positioned outside the perimeter of the training set in the chemical space , spanned by the first three principal components, which explain 88.6% of the variance in the original data, as demonstrated by Figure 7, suggesting that they serve as a good metric for the robustness evaluation of a predictive model.
Molecular distribution for those samples in the training set (filled circle), the test set (open triangle) and the outlier set (gray square) in the chemical space spanned by three principal components.
The prediction results of those molecules in the outlier set are listed in Table S1 and their associated statistical evaluations are summarized in Table 1. Hypo A, Hypo B and Hypo C yielded r2 values of 0.79, 0.70 and 0.84, respectively, in the outlier set, implying various performance decreases from the training set. Conversely, RMSE, ΔMax, MAE and s indicated that the performances of Hypo A, Hypo B and Hypo C increased from the training set to the outlier set because of lowered values of those parameters. However, this seemingly unusual characteristic for a predictive model can be realized by the fact that, of 11 molecules in the outlier set, the inhibition activities of 10 molecules are in the same log unit, viz. very close activities.
Similar to the observations found in the training set and test set, this PhE/SVM model performed better than any of pharmacophore models in the ensemble in the outlier set as indicated by those statistical parameters (Table 1) as well as the scatter plot of observed vs. predicted pEC50 values (Figure 8). Furthermore, the predictions by PhE/SVM are in extremely good agreement with observed values for all of molecules in the outlier set as manifested by the fact that the RMSE, ΔMax, MAE and s values are only 0.10, 0.13, 0.09 and 0.05, respectively, which are also smaller than their counterparts in the training set. The parameter r2 evaluated by PhE/SVM even increased from 0.89 in the training set to 0.96 in the outlier set. These statistical evaluations assert the fact that PhE/SVM is completely insensitive to the outliers, suggesting that it is a very robust predictive model as a result, which is of pivotal importance to practical applications.
Observed pEC50 vs. the pEC50 predicted by Hypo A, Hypo B, Hypo C and SVM model for those molecules in the outlier set. The solid line, dashed lines and dotted lines correspond to the SVM regression of the data, 95% confidence interval for the SVM regression and 95% confidence interval for the prediction, respectively.
The predictivity of generated PhE/SVM model was further evaluated by those validation requirements proposed by Golbraikh et al.  as well as Roy and Roy  in the training set, test set and outlier set. The results, summarized in Table 3, indicate that PhE/SVM not only yielded high statistical values but also met all validation requirements, suggesting that this predictive model is highly accurate and predictive. Furthermore, this PhE/SVM model can maintain similar performances regardless in the training set, test set and even outlier set as depicted by the little variations among different data set. As a result, it is plausible to expect, based on the facts mentioned above, that no substantial prediction errors will be generated when applied to structurally novel compounds.
It has been experimentally proven that P-gp has multiple binding sites . As a result, Ekins et al. produced four pharmacophore hypotheses, which consisted of different combinations of chemical features, based on different sets of samples . More importantly, the discrepancies in feature selections among these four models are consistent with the fact that Hypo A, Hypo B and Hypo C in the PhE also employed different chemical features, suggesting that different chemotypes of inhibitors can interact with P-gp using different chemical interactions, which completely agrees with the observation of Pajeva et al. , . Thus, only a group of fixed chemical features, viz. a single pharmacophore hypothesis, cannot fully take into account the promiscuous nature of P-gp.
Furthermore, those four pharmacophore models developed by Ekins et al. collectively consisted of HBD, HBA, HP and ring aromatic (RA) as compared with PhE/SVM, which was collectively composed of HBD, HBA and HP, indicating that the only qualitative difference between those 4 models and PhE/SVM is the absence of RA in the latter. Statistically, the lack of RA does not deteriorate the performance of PhE/SVM as compared with those four pharmacophore models. For instance, those four predictive models generated the r2 values of 0.77, 0.88, 0.86 and 0.76 in the training set, whereas PhE/SVM produced a value of 0.89, suggesting that the chemical feature RA is not a ncecssity to develope a predictive model. As a result, it is plausible to replace RA by HP, which can be manifested by the fact that the pharmacophore model developed by Palmeira et al. , which comprised one HBA and two RAs, predicted that the two RAs fitted onto the aromatic rings of propafenone, which, in turn, were depicted as hydrophobic by the predictive model proposed by Pajeva and Wiese . In fact, none of published predictive inhibition models enrolled the chemcial feature RA except those developed by Ekins et al. ,  and Palmeira et al. .
Furthermore, at least one HBA and one HP can always be found among all published pharmacophore hypotheses for P-gp inhibitors –, ,  except those models proposed by Ekins et al.  and Palmeira et al. . Collectively, PhE/SVM also consisted of the chemical features HBA, HBD and HP. Nevertheless, only Hypo B adopted the chemical feature HBA among those three models in the PhE. This seemingly paradox can be understood by the fact that one of predictive models developed by Ekins et al.  did not employ the chemical feature HBA, suggesting that not all inbitors interact with P-gp using HBA. In other words, it is not necessary to always take into account HBA. As a result, it is plausible to observe that not all of hypotheses in the ensemble selected the chemical feature HBA.
Langer et al. developed a pharmacophore hypothesis, which was composed of the chemical features (aromatic) HP, HBA and positive ionizable (PI) . Of 106 samples in the test set, whose experimental values were no larger than 207 µM, 34 molecules were projected as inactive since their predictive values were larger than 3,100,000 µM. These substantial discrepancies between the obseved values and predictions indicate their indiscriminations against these samples that was plausibly due to the lack of some key chemical features . Conversely, the PhE/SVM model, which was collectively comprised of the chemical features HBD, HBA and HP, yielded residuals of no more than 1 log unit for those same 34 molecules, asserting that PhE/SVM is a much more accurate model and those important features were completely taken into consideration. The most pronounced discrepancy between both theoretical models are resulted from the prediction of 24, which yielded residuals of 4.63 and 0.52 by the model derived by Langer et al. and PhE/SVM, respectively. Thus, it is presumable to attribute the qualitative differences between both theoretical models to the fact that Langer et al. enlisted the chemical feature PI without taking into account HBD, whereas PhE/SVM chose HBD over PI, suggesting that HBD plays a key role in inhibitor–P-gp interaction. The importance of HBD can be further manifested by 13, for example, whose hydroxyl group can be perfectly fitted to the chemical feature HBD in Hypo A, Hypo B and Hypo C as illustrated by Figure 9.
Pharmacophore models (A) Hypo A, (B) Hypo B and (C) Hypo C fitted to 13. The chemical features are described in Figure 2.
In addition, numerous studies have demonstrated the importance of HBD in determining the interaction between inhibitor and P-gp. For instance, Ekins et al.  and Pajeva et al. ,  recruited the chemical feature HBD to develop their pharmacophore hypotheses. Wang et al. , Zalloum and Taha  and Chen et al.  employed HBD related descriptor to construct their QSAR models; and even the CoMSIA model proposed by Labrie et el.  also used the field HBD. Accordingly, it is plausible to assume that the chemical feature HBD plays a critical role in determining the interaction between inhibitor and P-gp. Otherwise, any theoretical model may give rise to substantial prediction errors for some molecules.
P-gp inhibition is vital for drug metabolism and pharmacokinetics profiling since it can lead to adverse drug-drug interactions or even toxicity. A predictive model can be greatly valuable to drug discovery and development. Nevertheless, any in silico model that fails to take into account the promiscuous nature of P-gp cannot accurately model the interactions between structurally distinct inhibitors and P-pg. In this study, a quantitative predictive model, derived from a novel scheme by assembling a panel of pharmacophore hypothesis candidates to construct pharmacophore ensemble, which takes into consideration protein plasticity, and support vector machine, which generates a regression model, was developed to predict the P-gp inhibition. This developed PhE/SVM showed excellent prediction accuracy for those structurally diverse 31 and 88 molecules in the training set and test set, respectively, with excellent predictivity and statistical significance. It also executed extremely well when applied to those molecules in the outlier set, which were structurally dissimilar to those in the training set, as compared with any other conventional pharmacophore models, which adopted fixed selections of chemical features and can be only used to model molecules of specific chemical structures, substantially limiting their applicability as a result. Furthermore, the PhE/SVM model can elucidate the discrepancies among all published pharmacophore models, suggesting its superiority over the other theoretical models. Thus, it can be asserted that this PhE/SVM model can be adopted as an accurate and reliable predictive tool, even in the high throughput fashion, to facilitate drug discovery and development by designing drug candidates with better pharmacokinetic profile in terms of better absorption, higher bioavailability and more efficacy.
Selected compounds for this study, their names, SMILES strings, observed pEC50 values and predicted values by Hypo A, Hypo B, Hypo C and PhE/SVM, data partitions and references.
Parts of calculations were carried out at the National Center for High-Performance Computing, Taiwan. We thank Dr. G. H. Hakimelahi for reading the manuscript.
Conceived and designed the experiments: MKL. Performed the experiments: MKL HBC YHS. Analyzed the data: MKL HBC YHS. Contributed reagents/materials/analysis tools: MKL HBC YHS. Wrote the paper: MKL HBC YHS.
- 1. Schinkel AH, Jonker JW (2003) Mammalian drug efflux transporters of the ATP binding cassette (ABC) family: an overview. Adv Drug Deliver Rev 55: 3–29.
- 2. Fojo AT, Ueda K, Slamon DJ, Poplack DG, Gottesman MM, et al. (1987) Expression of a multidrug-resistance gene in human tumors and tissues. Proc Natl Acad Sci USA 84: 265–269.
- 3. Thiebaut F, Tsuruo T, Hamada H, Gottesman MM, Pastan I, et al. (1987) Cellular localization of the multidrug-resistance gene product P-glycoprotein in normal human tissues. Proc Natl Acad Sci USA 84: 7735–7738.
- 4. Ambudkar SV, Kimchi-Sarfaty C, Sauna ZE, Gottesman MM (2003) P-glycoprotein: from genomics to mechanism. Oncogene 22: 7468–7485.
- 5. Gottesman MM, Pastan I (2003) Biochemistry of Multidrug Resistance Mediated by the Multidrug Transporter. Ann Rev Biochem 62: 385–427.
- 6. Doppenschmitt S, Spahn-Langguth H, Regårdh CG, Langguth P (1999) Role of P-glycoprotein-mediated secretion in absorptive drug permeabiity: An approach using passive membrane permeability and affinity to P-glycoprotein. J PharmSci 88: 1067–1072.
- 7. Bansal T, Jaggi M, Khar RK, Talegaonkar S (2009) Status of Flavonols as P-Glycoprotein Inhibitors in Cancer Chemotherapy. Curr Cancer Ther Rev 5: 89–99.
- 8. Mistry P, Stewart AJ, Dangerfield W, Okiji S, Liddle C, et al. (2001) In Vitro and in Vivo Reversal of P-Glycoprotein-mediated Multidrug Resistance by a Novel Potent Modulator, XR9576. Cancer Res 61: 749–758.
- 9. Ambudkar S, Dey S, Hrycyna C, Ramachandra M, Pastan I, et al. (1999) Biochemical, cellular, and pharmacological aspects of the multidrug transporter. Annu Rev Pharmacol Toxicol 361–398.
- 10. Kim RB, Fromm MF, Wandel C, Leake B, Wood AJ, et al. (1998) The drug transporter P-glycoprotein limits oral absorption and brain entry of HIV-1 protease inhibitors. J Clin Invest 101: 289–294.
- 11. Löscher W, Potschka H (2005) Drug resistance in brain diseases and the role of drug efflux transporters. Nat Rev Neurosci 6: 591–602.
- 12. Jansen G, Scheper R, Dijkmans B (2003) Multidrug resistance proteins in rheumatoid arthritis, role in disease-modifying antirheumatic drug efficacy and inflammatory processes: an overview. Scand J Rheumatol 32: 325–336.
- 13. Szakacs G, Paterson JK, Ludwig JA, Booth-Genthe C, Gottesman MM (2006) Targeting multidrug resistance in cancer. Nat Rev Drug Discov 5: 219–234.
- 14. Solit DB, Rosen N (2006) Hsp90: A Novel Target for Cancer Therapy. Curr Top Med Chem 6: 1205–1214.
- 15. Biamonte MA, Van de Water R, Arndt JW, Scannevin RH, Perret D, et al. (2009) Heat Shock Protein 90: Inhibitors in Clinical Trials. J Med Chem 53: 3–17.
- 16. Taldone T, Gozman A, Maharaj R, Chiosis G (2008) Targeting Hsp90: small-molecule inhibitors and their clinical development. Curr Opin Pharmacol 8: 370–374.
- 17. Kim YS, Alarcon SV, Lee S, Lee MJ, Giaccone G, et al. (2009) Update on Hsp90 Inhibitors in Clinical Trial. Curr Top Med Chem 9: 1479–1492.
- 18. Coley HM (2010) Overcoming Multidrug Resistance in Cancer: Clinical Studies of P-Glycoprotein Inhibitors. In: Zhou J, editor. Multi-Drug Resistance in Cancer. New York: Humana Press. pp. 341–358.
- 19. Shukla S, Wu C-P, Ambudkar SV (2008) Development of inhibitors of ATP-binding cassette drug transporters – present status and challenges. Expert Opin Drug Metab Toxicol 4: 205–223.
- 20. Leonard GD, Fojo T, Bates SE (2003) The Role of ABC Transporters in Clinical Practice. Oncologist 8: 411–424.
- 21. Lin JH (2003) Drug-drug interaction mediated by inhibition and induction of P-glycoprotein. Adv Drug Deliv Rev 55: 53–81.
- 22. Greiner B, Eichelbaum M, Fritz P, Kreichgauer H-P, von Richter O, et al. (1999) The role of intestinal P-glycoprotein in the interaction of digoxin and rifampin. J Clin Invest 104: 147–153.
- 23. Crowley E, McDevitt CA, Callaghan R (2010) Generating Inhibitors of P-Glycoprotein: Where to, Now? In: Zhou J, editor. Multi-Drug Resistance in Cancer. New York: Humana Press. pp. 405–432.
- 24. Ekins S, Mestres J, Testa B (2007) In silico pharmacology for drug discovery: applications to targets and beyond. Br J Pharmacol 152: 21–37.
- 25. Chang C, Ekins S (2006) Pharmacophores for Human ADME/Tox-related Proteins. In: Langer T, Hoffmann RD, editors. Pharmacophores and Pharmacophore Searches. Weinheim, Germany: Wiley. pp. 299–324.
- 26. Zhou H, Wu S, Zhai S, Liu A, Sun Y, et al. (2008) Design, synthesis, cytoselective toxicity, structure-activity relationships, and pharmacophore of thiazolidinone derivatives targeting drug-resistant lung cancer cells. J Med Chem 51: 1242–1251.
- 27. Chang C, Bahadduri PM, Polli JE, Swaan PW, Ekins S (2006) Rapid Identification of P-glycoprotein Substrates and Inhibitors. Drug Metab Dispos 34: 1976–1984.
- 28. Langer T, Eder M, Hoffmann RD, Chiba P, Ecker GF (2004) Lead identification for modulators of multidrug resistance based on in silico screening with a pharmacophoric feature model. Arch Pharm 337: 317–327.
- 29. Pajeva IK, Wiese M (2002) Pharmacophore model of drugs involved in P-glycoprotein multidrug resistance: explanation of structural variety (hypothesis). J Med Chem 45: 5671–5686.
- 30. Ekins S, Kim RB, Leake BF, Dantzig AH, Schuetz EG, et al. (2002) Three-dimensional quantitative structure-activity relationships of inhibitors of P-glycoprotein. Mol Pharmacol 61: 964–973.
- 31. Ekins S, Kim RB, Leake BF, Dantzig AH, Schuetz EG, et al. (2002) Application of three-dimensional quantitative structure-activity relationships of P-glycoprotein inhibitors and substrates. Mol Pharmacol 61: 974–981.
- 32. Pajeva I, Globisch C, Fleischer R, Tsakovska I, Wiese M (2005) Molecular Modeling of P-Glycoprotein and Related Drugs. Med Chem Res 14: 106–117.
- 33. Palmeira A, Rodrigues F, Sousa E, Pinto M, Vasconcelos M, et al. (2010) Pharmacophore-Based Screening as a Clue for the Discovery of New P-Glycoprotein Inhibitors. In: Rocha M, Riverola F, Shatkay H, Corchado J, editors. Advances in Bioinformatics. Berlin: Springer Berlin/Heidelberg. pp. 175–180.
- 34. Pleban K, Kaiser D, Kopp S, Peer M, Chiba P, et al. (2005) Targeting drug-efflux pumps—a pharmacoinformatic approach. Acta Biochim Pol 52: 737–740.
- 35. Ekins S (2004) Predicting undesirable drug interactions with promiscuous proteins in silico. Drug Discov Today 9: 276–285.
- 36. Loo TW, Bartlett MC, Clarke DM (2003) Substrate-induced conformational changes in the transmembrane segments of human P-glycoprotein. Direct evidence for the substrate-induced fit mechanism for drug binding. J Biol Chem 278: 13603–13606.
- 37. Bahadduri P, Polli J, Swaan P, Ekins S (2010) Targeting Drug Transporters – Combining In Silico and In Vitro Approaches to Predict In Vivo. In: Yan Q, editor. Membrane Transporters in Drug Discovery and Development: Methods and Protocols. New York: Humana Press. pp. 65–103.
- 38. Ward A, Reyes CL, Yu J, Roth CB, Chang G (2007) Flexibility in the ABC transporter MsbA: Alternating access with a twist. Proc Natl Acad Sci 104: 19005–19010.
- 39. Stockner T, de Vries SJ, Bonvin AM, Ecker GF, Chiba P (2009) Data-driven homology modelling of P-glycoprotein in the ATP-bound state indicates flexibility of the transmembrane domains. FEBS J 276: 964–972.
- 40. Ravna A, Sylte I, Sager G (2009) Binding site of ABC transporter homology models confirmed by ABCB1 crystal structure. Theoretical Biology and Medical Modelling 6: 20.
- 41. Aller SG, Yu J, Ward A, Weng Y, Chittaboina S, et al. (2009) Structure of P-Glycoprotein Reveals a Molecular Basis for Poly-Specific Drug Binding. Science 323: 1718–1722.
- 42. Gutmann DAP, Ward A, Urbatsch IL, Chang G, van Veen HW (2010) Understanding polyspecificity of multidrug ABC transporters: closing in on the gaps in ABCB1. Trends Biochem Sci 35: 36–42.
- 43. Ecker GF (2010) QSAR Studies on ABC Transporter – How to Deal with Polyspecificity. In: Ecker G, Chiba P, editors. Transporters as Drug Carriers: Structure, Function, Substrates. Weinheim, Germany: Wiley-VCH Verlag GmbH & Co. KGaA. pp. 195–214.
- 44. Varma MVS, Ashokraj Y, Dey CS, Panchagnula R (2003) P-glycoprotein inhibitors and their screening: a perspective from bioavailability enhancement. Pharmacol Res 48: 347–359.
- 45. Globisch C, Pajeva IK, Wiese M (2008) Identification of Putative Binding Sites of P-glycoprotein Based on its Homology Model. Chem Med Chem 3: 280–295.
- 46. Penzotti JE, Lamb ML, Evensen E, Grootenhuis PDJ (2002) A Computational Ensemble Pharmacophore Model for Identifying Substrates of P-Glycoprotein. J Med Chem 45: 1737–1740.
- 47. Li W-X, Li L, Eksterowicz J, Ling XB, Cardozo M (2007) Significance analysis and multiple pharmacophore models for differentiating P-glycoprotein substrates. J Chem Inf Model 47: 2429–2438.
- 48. Leong MK (2007) A Novel Approach Using Pharmacophore Ensemble/Support Vector Machine (PhE/SVM) for Prediction of hERG Liability. Chem Res Toxicol 20: 217–226.
- 49. Chen C-N, Shih Y-H, Ding Y-L, Leong MK (2011) Predicting Activation of the Promiscuous Human Pregnane X Receptor by Pharmacophore Ensemble/Support Vector Machine Approach. Chem Res Toxicol 24: 1765–1778.
- 50. Leong MK, Chen Y-M, Chen H-B, Chen P-H (2009) Development of a New Predictive Model for Interactions with Human Cytochrome P450 2A6 Using Pharmacophore Ensemble/Support Vector Machine (PhE/SVM) Approach. Pharm Res 26: 987–1000.
- 51. Leong MK, Chen T-H (2008) Prediction of cytochrome P450 2B6-substrate interactions using pharmacophore ensemble/support vector machine (PhE/SVM) approach. Med Chem 4: 396–406.
- 52. Chiba P, Holzer W, Landau M, Bechmann G, Lorenz K, et al. (1998) Substituted 4-Acylpyrazoles and 4-Acylpyrazolones: Synthesis and Multidrug Resistance-Modulating Activity. J Med Chem 41: 4001–4011.
- 53. Klein C, Kaiser D, Kopp S, Chiba P, Ecker GF (2002) Similarity based SAR (SIBAR) as tool for early ADME profiling. J Comput-Aided Mol Des 16: 785–793.
- 54. Hiessböck R, Wolf C, Richter E, Hitzler M, Chiba P, et al. (1999) Synthesis and in Vitro Multidrug Resistance Modulating Activity of a Series of Dihydrobenzopyrans and Tetrahydroquinolines. J Med Chem 42: 1921–1926.
- 55. Foloppe N, Chen I-J (2009) Conformational Sampling and Energetics of Drug-Like Molecules. Curr Med Chem 16: 3381–3413.
- 56. Chang G, Guida WC, Still WC (1989) An internal-coordinate Monte Carlo method for searching conformational space. J Am Chem Soc 111: 4379–4386.
- 57. Kolossvary I, Guida WC (1996) Low mode search. An efficient, automated computational method for conformational analysis: Application to cyclic and acyclic alkanes and cyclic peptides. J Am Chem Soc 118: 5011–5019.
- 58. Halgren TA (1996) Merck molecular force field. I. Basis, form, scope, parameterization, and performance of MMFF94. J Comput Chem 17: 490–519.
- 59. Still WC, Tempczyk A, Hawley RC, Hendrickson T (1990) Semianalytical treatment of solvation for molecular mechanics and dynamics. J Am Chem Soc 112: 6127–6129.
- 60. Zou J, Xie H-Z, Yang S-Y, Chen J-J, Ren J-X, et al. (2008) Towards more accurate pharmacophore modeling: Multicomplex-based comprehensive pharmacophore map and most-frequent-feature pharmacophore model of CDK2. J Mol Graph Model 27: 430–438.
- 61. Sprague PW (1995) Automated chemical hypothesis generation and database searching with Catalyst. Perspect Drug Discovery Des 3: 1–20.
- 62. Li H, Sutter J, Hoffmann R (2000) HypoGen: An Automated System for Generating 3D Predictive Pharmacophore Models. In: Güner OF, editor. Pharmacophore Perception, Development, and Use in Drug Design. La Jolla, California: International University Line. pp. 171–189.
- 63. Labrie P, Maddaford SP, Lacroix J, Catalano C, Lee DKH, et al. (2006) In vitro activity of novel dual action MDR anthranilamide modulators with inhibitory activity at CYP-450. Bioorg Med Chem 14: 7972–7987.
- 64. Kurogi Y, Güner OF (2001) Pharmacophore Modeling and Three-dimensional Database Searching for Drug Design Using Catalyst. Curr Med Chem 8: 1035–1055.
- 65. Evans DA, Doman TN, Thorner DA, Bodkin MJ (2007) 3D QSAR methods: Phase and Catalyst compared. J Chem Inf Model 47: 1248–1257.
- 66. Leong MK, Chen Y-M, Chen T-H (2009) Prediction of Human Cytochrome P450 2B6-Substrate Interactions Using Hierarchical Support Vector Regression Approach. J Comput Chem 30: 1899–1909.
- 67. Golbraikh A, Shen M, Xiao Z, Xiao Y-D, Lee K-H, et al. (2003) Rational selection of training and test sets for the development of validated QSAR models. J Comput-Aided Mol Des 17: 241–253.
- 68. Benfenati E, Chrétien JR, Giuseppina Gini, Piclin N, Pintore M, et al. (2007) Validation of the models. In: Benfenati E, editor. Quantitative Structure-Activity Relationships (QSAR) for Pesticide Regulatory Purposes. Amsterdam: Elsevier. pp. 185–199.
- 69. Roy PP, Roy K (2008) On Some Aspects of Variable Selection for Partial Least Squares Regression Models. QSAR Comb Sci 27: 302–313.
- 70. Gramatica P, Giani E, Papa E (2007) Statistical external validation and consensus modeling: A QSPR case study for Koc prediction. J Mol Graph Model 25: 755–766.
- 71. Martin C, Berridge G, Higgins CF, Mistry P, Charlton P, et al. (2000) Communication between Multiple Drug Binding Sites on P-glycoprotein. Mol Pharmacol 58: 624–632.
- 72. Wang Y-H, Li Y, Yang S-L, Yang L (2005) An in silico approach for screening flavonoids as P-glycoprotein inhibitors based on a Bayesian-regularized neural network. J Comput-Aided Mol Des 19: 137–147.
- 73. Zalloum HM, Taha MO (2008) Development of predictive in silico model for cyclosporine- and aureobasidin-based P-glycoprotein inhibitors employing receptor surface analysis. J Mol Graph Model 27: 439–451.
- 74. Chen L, Li Y, Zhao Q, Peng H, Hou T (2011) ADME Evaluation in Drug Discovery. 10. Predictions of P-Glycoprotein Inhibitors using Recursive Partitioning and Naïve Bayesian Classification Techniques. Mol Pharmaceutics 8: 889–900.
- 75. Labrie P, Maddaford SP, Fortin S, Rakhit S, Kotra LP, et al. (2006) A comparative molecular field analysis (CoMFA) and comparative molecular similarity indices analysis (CoMSIA) of anthranilamide derivatives that are multidrug resistance modulators. J Med Chem 49: 7646–7660.