A Novel Approach for Predicting P-glycoprotein (ABCB1) Inhibition Using Molecular Interaction Fields

Abstract

P-glycoprotein (Pgp or ABCB1) is an ABC transporter protein involved in intestinal absorption, drug metabolism and brain penetration, and its inhibition can seriously alter a drug's bioavailability and safety. In addition, inhibitors of Pgp can be used to overcome multidrug resistance. Given this dual-purpose, reliable in silico procedures to predict Pgp inhibition are of great interest. A large and accurate literature collection yielded more than 1200 structures; a model was then constructed using various MIF-based technologies, considering pharmacophoric features and those physico-chemical properties related to membrane partitioning. High accuracy was demonstrated internally, with two different validation sets, and moreover using a number of molecules, for which Pgp inhibition was not experimentally available but was evaluated `in-house'. All the validations confirmed the robustness of the model and its suitability to help medicinal chemists in drug discovery. The information derived from the model was rationalized as a pharmacophore for competitive Pgp inhibition.

Introduction

Minimizing failures in early phases is one of the main strategies in current drug discovery. While in the period between 1964–1985 poor pharmacokinetic properties were the major reason for drug failures, in the last two decades safety, together with lack of efficacy, have become a main concern. In particular, human adverse drug reactions (ADRs) have emerged as the principal reason for drugs withdrawal from the market over the past 20 years.¹ Rarely occurring ADRs may explain why potentially toxic effects of drugs were not detected during clinical trials. Rationalization of these failures led to the identification of a number of antitargets, namely those targets which, upon interaction with therapeutic drugs, may result in severe ADRs.² ABCB1, also known as P-glycoprotein, Pgp or MDR1, is a membrane protein member of the ATP-binding cassette (ABC) transporters superfamily. Together with hERG channel and CYP3A4, it is probably the most widely studied antitarget.

P-glycoprotein is expressed in a variety of human tissues as defense against xenobiotics. It uses energy derived by ATP hydrolysis to translocate its ligands out of the cell against the concentration gradient. Pgp is probably the most promiscuous efflux transporter, since it recognizes a number of structurally different and apparently unrelated xenobiotics; notably, many of them are also CYP3A4 substrates. CYP3A4 and ABCB1 are often expressed in the same tissues, hence for common substrates the amount of efflux determines the exposure to metabolism.³ This interplay affects bioavailability of drugs co-administered with a Pgp inhibitor or inducer.^4,5

Pgp is also an interesting target in oncology,⁶ since multidrug resistance is often associated with its overexpression. Therefore, potent selective Pgp inhibitors have been rationalized as adjuvant therapy when co-administered with anti-cancer drugs. Until now, a number of candidates failed clinical trials due to poor selectivity. In particular, first generation chemosensitizers, generally drugs known to be active toward other targets, were ineffective at non-toxic concentrations, while second generation chemosensitizers often failed because of simultaneous Pgp and CYP3A4 inhibition.⁷

Pgp awareness should be routinely included in the early phases of drug discovery,⁸ given its duality as target and antitarget.⁹ Reliable in vitro assays to evaluate the Pgp inhibition capability of new drug entities are costly and time demanding, so the integration of in silico and in vitro procedures can help to minimize the costs. For this reason a number of in silico models for recognition of Pgp substrates and inhibitors have been proposed in recent years. Lack of a high-resolution crystal structure for human Pgp, together with the high flexibility of ABC transporters, justify the prevalence of ligand-based models.¹⁰ Statistics and information gained from these models were recently reviewed in extenso.^11,12 The reviewed Pgp inhibition models generally agreed on the utility of pharmacophoric descriptors, and consistently identified the importance of a hydrogen bond (HB) acceptor coupled with some hydrophobic regions (between two and four). Although these models shared good interpretability, they demonstrated diminished performance when tested against non-local external validation sets. On the other hand, classification models using non-pharmacophoric description often showed better predictive power for Pgp substrate recognition, but were rarely used to discriminate inhibitors from non-inhibitors.

It is preferable for in silico models to rely upon on an extensive data collection that allows an appropriate chemical space coverage, combined with appropriate molecular descriptors. In this work a thorough literature analysis yielded a training set of 772 molecules and two validation sets, composed of molecules taken either from the same references of the training set (with inherent chemical space bias), or from articles not used for the training set collection (i.e. different chemotypes). In addition, different classes of molecular description were evaluated, in order to account for non-specific factors such as water solubility and membrane partitioning and for more specific pharmacophoric features responsible for ligand-protein interactions.

In particular, molecules described using GRID Molecular Interaction Field (MIF)^13–15 approaches resulted in a “Composite model” for Pgp inhibition, based on an intuitive pipeline of preceding “local models” developed for competitive or non-competitive binding. VolSurf+ descriptors^16–18 were used to model physicochemical properties, whereas the pharmacophoric method FLAP^19–21 was employed to superimpose molecules and evaluate their MIF similarity. With VolSurf+, a number of important parameters were detected for Pgp inhibition, while FLAP identified the most important pharmacophoric features around the optimal molecular shape. The emergent model can be used to predict Pgp inhibition and to guide compound design to modulate its impact within series of molecules.

Results

Data Collection

The collected data (available as Supporting Information) refers to the inhibition of transport of an ABCB1 probe substrate in a cell line expressing Pgp. For compound categories we observed the definitions of Rautio et al., who tested some molecules with different protocols and derived qualitative rules to classify Pgp inhibitors.²² Thus, inhibitors having IC₅₀ lower than 15 μM may be involved in drug-drug interactions, whereas those with an IC₅₀ higher than 100 μM are considered not inhibitors. Molecules having an IC₅₀ of between 15 and 100 μM are classified as weak inhibitors. Different protocols revealed that some inhibitors (IC₅₀ lower than 15 μM) and non-inhibitors fall in the “grey region” of weak inhibitors when experimental conditions were altered.²² These compounds were excluded from our set. On the contrary, a “two-class jump” (from inhibitor to non- inhibitor) was rarely observed, e.g. for 9-hydroxy-Risperidone.^23,24

To enlarge our dataset we extended the qualitative classification to all those molecules for which activity was reported as percentage of inhibition. Thresholds values for percentage of inhibition were defined through pairwise comparison based on those structures for which both IC₅₀ and percentage were available (see Table 1).

Table 1.

Criteria used to classify molecules either as Pgp non-inhibitor or inhibitor.

	IC50 (μM)	% Inhibitiona	% Inhibitionb
Non-inhibitor	≥ 100	< 10	< 12
Inhibitor	≤ 15	> 25	> 30

^a% Inhibition respect to not transfected cell line or maximum inhibition.³⁸

^bFor the particular case of PDSP different thresholds were defined.⁶⁵

Molecules with molecular weight >700 were not used to avoid inappropriate coverage of the chemical space. Rules for categorization were applied to data from 44 publications, resulting in a set of 857 molecules, which were then divided into a “Training Set” (N=772) and an “Internal Validation Set” (N=85), as illustrated in Figure 1.^22–65 The Internal Validation Set was extracted from a Principal Component Analysis (PCA) computed in Volsurf+ using the Most Descriptive Compounds (MDC) algorithm for subset selection. This set was used to evaluate predictive performance of the model against known near-neighbor molecules (biased predictions).

Figure 1.

Data set selection. References and the corresponding molecules are differently coloured if they belong to the Training set and Internal Validation sets (blue) or External Validation set (magenta). In addition, for each set the division between Pgp inhibitors (red) and non-inhibitors (green) is shown with the horizontal bars. All data are available as Supporting Information.

A second “External Validation Set” was extracted from Wombat,⁶⁶ Chembl⁶⁷ and Klopman⁶⁸ data sets, excluding the molecules already present in our collection. Since Klopman et al. tested molecules for Multidrug Resistance and not specifically for Pgp inhibition, only those [not reversing] multidrug resistance (R-FOLD=1) were used. At the end, the External Validation Set was composed of N=418 molecules from 17 different references.^68–84 Projecting these molecules in the chemical space of the Training set (using Volsurf+ descriptors and Principal Component Analysis) confirmed that the chemical space of the two sets was only partially overlapping (projections are given as Supporting Information). This set was used to evaluate predictive performance of the model against unknown chemotypes (non-local, or unbiased predictions).

The Model

Molecules were imported into VolSurf+ as SMILES (mostly taken from PubChem⁸⁵ and Wombat⁶⁶); some were manually input via the Volsurf+ chemical sketcher.¹⁶ Molecular descriptors obtained with VolSurf+ and the pharmacophoric description obtained with FLAP¹⁹ were generated for the neutral species, since Ecker confirmed the role of potentially positive (basic) nitrogen as HB acceptor.⁸⁶

Both methods use GRID molecular interaction fields (MIF),^14,15 extracting and processing the information contained therein in different and complementary ways. GRID MIFs have been applied to many areas of drug discovery,^15,87,88 including pK_a and tautomer modeling,^89,90 structure-based drug design,⁹¹ scaffold-hopping,^92–94 3D-QSAR,^95–98 docking,^98,99 ADME and pharmacokinetic modeling,^17,18,100 and metabolism prediction.^101–102

Multivariate methods such as PCA or PLS are the proper way to elaborate pattern recognition or to study quantitative structure-activity relationships, respectively, when the molecules are described by means of Volsurf+ descriptors. A single conformation is typically used for modeling each molecule; however, due to the nature of descriptors, VolSurf+ is mostly conformation-independent. VolSurf+ molecular descriptors refer to molecular size and shape, to size and shape of hydrophilic and hydrophobic regions, and to the balance between them. MIFs from the OH2 and DRY GRID probes are used to define the hydrophilic and hydrophobic regions, respectively; however, the specification in the molecule of HB donor/acceptor character of polar regions, available using the GRID probes O and N1, can be significant in certain cases. VolSurf+ descriptors have been presented in detail elsewhere.^17,18

In the FLAP approach, originally conceived for^19–21 and successfully applied to virtual screening,^103–107 molecules are stored in a database together with their MIFs, which are condensed into discrete points and exhaustively combined to produce quadruplets, or 4-point pharmacophores. Molecules are compared pairwise, with the quadruplets of each database molecule (ligand) being compared with the quadruplets of the reference structure (molecular template). When a search is performed, all of the quadruplets from the ligand are searched against all of the quadruplets in the template; quadruplets that match are then used to overlay the ligand onto the template. For each alignment, MIF similarity scores are used as a measure of the superimposition between the ligand MIF and the template MIF. A score can refer to the best alignment of two molecules according to a single probe, but two or more probes can be handled as well; in this case, the product of the probe scores is used to select the orientation that simultaneously represents the best MIF alignment of the given probes. The best similarity value for each MIF type produced by a single conformer is retained for each molecule.

Many flavones and steroids, which are non-competitive Pgp inhibitors,¹⁰⁸ were found as outliers by both VolSurf+ and FLAP preliminary models. Thus, 94 molecules with at least one of these chemotypes (flavone-like and steroid-like) were identified among the Training Set and were used to compute a “local model” with the same methodologies used for the Training Set. Attempts to define a pharmacophore model for such a set of compounds (using the same procedure described below for the larger training set) were not as satisfactory as expected; the presence of two different suspect substructures is likely to affect the pharmacophore modeling (with FLAP, whereas it was possible to obtain a model with VolSurf+). In particular, the VolSurf+ model was based on the first two latent variables (LVs), indicating a determinant role for the descriptors of hydrophobicity and permeability (D3, LogPn-oct), which are directly related to Pgp inhibition, and an inverse correlation with aqueous solubility (SOLY).

The remaining 772 Training Set molecules (not containing flavones and steroid chemotypes) were used to derive a Volsurf+ based Partial Least Square Discriminant (PLS-D) model. The model was based on five LVs (Q² and R² are given as Supporting Information). In particular, HSA, LogPn-oct and descriptors of size (such as molecular surface), were relevant for the first LV, whereas the second and the third reported the importance of flexibility (FLEX), hydrophilic volume (W1), the pharmacophoric descriptor DRDRAC, which is the area between two hydrophobic atoms and one HB acceptor atom (second LV) and the “integy moment” IW4, which refers to the localisation of hydrophilic regions (third LV). Finally, the fourth and the fifth LVs suggested a determinant role of the charged state (L1lgS in the fourth LV, NCC in the fifth LV, %FU4–10 in both) and hydrophilic capacity factors (in the fourth LV).

The same Training Set molecules were used as input for a Linear Discriminant Analysis (LDA) model for Pgp inhibition with FLAP. The automated procedure available in FLAP selected the chemosensitizer Biricodar among twenty potent Pgp inhibitors candidate templates, and a combination of the GRID probes H, DRY, O and N1 was selected to compute the optimal LDA model. Once a predictive model was computed based on the fields derived by all the atoms of the template, the FLAP module for the selective pharmacophore extraction was applied. In this procedure, the software selects a number of defined features that are important for the LDA model. This model uses only the fields of the selected atoms, together with the shape of the template. Thus, a pharmacophore composed of one HB acceptor, two large hydrophobic regions and molecular shape was created for the competitive binding. The model was based on the FLAP scores H, H*N1, and DRY. Surprisingly, the probe shape (H) was the most relevant, contributing similiarity information to the model alone, and in combination with the N1 probe.

All these models were linked into a “composite model”. The pipeline is briefly reported below and schematized in Figure 3. As a first step, when the substructure search reveals the presence of suspect chemotypes, non-competitive binding is evaluated and molecules are classified either as inhibitors or passed further to the competitive binding module. Thus, the combined VolSurf+ and FLAP models evaluate molecules that are less likely to act as non-competitive inhibitors, which are then classified as Pgp inhibitors or non-inhibitors.

Figure 3.

Flow chart of the Composite Model.

The use of accuracy, sensitivity and specificity is commonly accepted to evaluate the capabilities and performance of classification models. In our classification system, sensitivity measures how a model is able to classify inhibitors, specificity measures how a model is able to classify non-inhibitors, whereas accuracy measures how a model is able to classify all the compounds, both inhibitors and non-inhibitors. Table 2 reports the performances of the model in the Training set, the Internal Validation set and the External Validation set.

Table 2.

Model performance in the different data sets.

	Accuracya	Sensitivitya	Specificitya	Unweighted Cohen's kappaa
Training	0.88	0.84	0.91	0.75
Internal Validation	0.85	0.82	0.87	0.69
External Validation	0.86	0.90	0.80	0.70

^aPercent values refer to the number of correct predictions with respect to the total number of molecules (N=772). Since Pgp inhibition is the target of the model, sensitivity refers to the prediction of inhibitors and specificity of non-inhibitors. All the statistict were derived using the True Positives, True Negatives, False Positive and False Negatives as reported in the Supporting Information.

A final “blind” validation set was composed of a set of eight compounds picked among a data set of about 2,700 drugs, from an in-house historical collection (DRUGS database). All of the compounds already present in our Pgp collection were excluded, with the only exception of Alprenolol (well known Pgp non-inhibitor) and Cyclosporine A (well known Pgp inhibitor) which were included as a negative and positive control, respectively. The remaining molecules of the large data set were projected onto the models and classified either as Pgp inhibitors or non-inhibitors. A snapshot of the prediction is reported in Figure 4: interestingly, 21.5% of the molecules were predicted as potential Pgp inhibitors, which reflects the known high promiscuity of Pgp. Both methods contributed to the prediction of Pgp inhibitors, however Volsurf+ was more selective in classifying compounds as non-inhibitors.

Figure 4.

Schema of the prediction of the DRUGS data set of drugs (about 2,700 molecules).

To make a blind selection, a random choice of two and six molecules was performed among those predicted as non-inhibitors or inhibitors, respectively. These eight molecules, together with Cyclosporine A (positive control) and Alprenolol (negative control) were analyzed experimentally to evaluate whether they inhibit human Pgp. The ability of these compounds to inhibit Pgp-mediated cell efflux of R123 in mouse T lymphoma L5178 MDR1 cells was studied. Alprenolol (control), Ticarcillin and Nateglinide were not effective, while all the other molecules tested are effective. In particular, Aprindine and Ziprasidone are weak inhibitors, Sertindole was characterized by α_max of 0.74 at the concentration of 1×10⁻⁴ M (apparent IC₅₀ value of 6.5×10⁻⁶ M) whereas Ebastine and Aripiprazole turned out to be efficacious Pgp inhibitors, with α_max values close to 1 and IC₅₀ value of 1.5×10⁻⁶ and 7.8×10⁻⁷ M, respectively. Both these compounds were found to be almost equipotent to Cyclosporine A.

Comparison with Other Methods

The Tanimoto similarities to the Pgp inhibitors Biricodar, Zosuquidar, Verapamil and Elacridar were calculated for each molecule in the Dataset using the software Open Babel.¹⁰⁹ The resulting scores were processed using the k Nearest Neighbor algorithm implemented in the open source data-mining software Orange Canvas.¹¹⁰ The euclidean distance was used as a distance metric and the k number of neighbors selected was 7. Parameters were selected on the basis of their performance on the training set and in internal validation. Predictions in leave-one-out on the Training set reached a classification accuracy of 0.76. The accuracy of the Internal Validation set was 0.78, while for the External Validation set the accuracy of the 2D method dropped to 0.59. Finally, out of the blind set the only Ebastine (Pgp inhibitor) and Ziprasidone (weak inhibitor) were predicted as Pgp inhibitors, whereas all the remaining molecules were predicted to be Pgp non-inhibitors.

Discussion

A novel approach for predicting P-glycoprotein (ABCB1) inhibition using GRID molecular interaction fields is reported. Instead of using one of the current literature collections available, an exhaustive literature search was performed, based on an accurate data analysis and on a straightforward way to categorize molecules as inhibitors and or non-inhibitors. More than 1,200 molecules were collected and structured into three data sets: the Training set and the Internal Validation set were extracted from the same 44 references (therefore sharing the same chemical space), whereas the External Validation set was derived from an additional 17 references.

The second step of the work focused on molecular description. Most in silico procedures for Pgp inhibitors recognition proposed in the past exclusively employed pharmacophoric descriptors, without considering descriptors typically related to ADME properties (e.g. membrane permeability). These models generally agree on the importance of HB acceptors and aromatic or hydrophobic features as a driving force for Pgp inhibition. The role of nitrogen as a HB acceptor has been experimentally confirmed, suggesting that molecules should be modeled in their neutral form.⁸⁶ Unfortunately, even when exhibiting good interpretability, pharmacophore models rarely reached high accuracy when applied to large datasets with non-local series. This suggests that a pharmacophore for Pgp inhibition is a necessary but not sufficient condition. Hence VolSurf+ molecular descriptors, optimized to model pharmacokinetic properties, were used in tandem with a pharmacophoric method (FLAP).

In a recent review Leach asserts that pharmacophoric hypotheses arise from the process of elucidation, defined as “the identification from a set of active molecules and their biological activities of the key common features and their relative orientations”; as Leach emphasizes, the process is dependent on the alignment of active molecules, that is the key step.¹¹¹ In this sense FLAP can be considered as pharmacophore-based, since all Training Set molecules are superimposed to a template's conformation and then GRID MIF similarity is calculated. Among the resulting scores, only a reduced number are subsequently employed in a Linear Discriminant Analysis. These are considered by FLAP the crucial features for the biological activity.

Using FLAP to define the pharmacophore it is possible to localize, orientate and confine the features supposed to be biologically essential. Interestingly, all the inhibitors in our data set have at least one HB acceptor atom, while there are some without HB donors. Particularly significant is the example of the tertiary amine Amitriptyline, an inhibitor with only one heteroatom,⁴² which leads to the minimalistic hypothesis of a pharmacophore with a unique HB acceptor atom, one or two hydrophobic and/or aromatic features and no HB donor atoms.

Based on the FLAP alignment, we hypothesize that an optimal shape for Pgp inhibitors may exist. Molecules in a conformation that sufficiently fits the optimal shape can be classified as Pgp inhibitors when they also have hydrophobic and HB acceptor characteristics (i.e., MIFs) that have a good similarity to that of the template's. The pharmacophore, reported in Figure 6, can be summarized as an ensemble of cavity-shape plus one hydrogen bond acceptor atom and one large hydrophobic region, together consistent with the minimalistic hypothesis reported above. It is interesting to note than almost 90% of all the inhibitors, small medium and large, fit this simple pharmacophoric hypotesis.

Figure 6.

FLAP pharmacophore hypothesis found using the FLAP-Pharm Procedure. Hydrophobic regions (where a favourable hydrophobic interaction might occur) are represented by the green contour, while HB acceptor regions are shown as red contours. The shape is coloured in pale grey. Buspirone (A) is superposed to the pharmacophore for reference. The buspirone ring on the right is probably exposed toward solvent. Other inhibitors (B=Tolafentrine; C=Simvastatin; D=Clemastine; E=Desmethylazelastine; F=D703; G=Noscapine) are reported to show the degree of complementarity with the proposed pharmacophore.

Notably, the use of molecular shape in the FLAP-derived pharmacophore not only fits, but fulfills the IUPAC definition: “A pharmacophore is the ensemble of steric and electronic features that is necessary to ensure the optimal supramolecular interactions with a specific biological target structure and to trigger (or to block) its biological response”.¹¹²

The precedent use of VolSurf+ descriptors, in agreement with the hypothesis of an important synergistic effect of pharmacokinetic and pharmacodynamic properties, produced a robust model. Descriptors of size, hydrophobic surface area, flexibility and LogPn-oct emerged as strongly significant (see Figure 7). In particular, while molecular surface, flexibility and hydrophobic surface might approximately encode characteristics captured by the FLAP model, LogPn-oct is a well-established membrane permeability descriptor. In fact a certain degree of correlation emerges in the work of Polli et al., where median and minimum permeability are sensibly higher for inhibitors respect to non-inhibitors.³⁸

Figure 7.

Distribution of some VolSurf+ descriptors for Pgp non-inhibitors (grey) and inhibitors (black). Descriptors reported are molecular surface (S), flexibility (FLEX), hydrophobic surface area (HSA) and coefficient of partitioning in octanol/water (LogPn-oct).

More than 90% of the molecules with a computed LogPn-oct lower than 2 are non-inhibitors (Figure 7). Descriptors of the charged state (e.g. percentage of neutral species at different pH) affected predictions in an opposite manner at the 3^th and at the 4^th latent variables. This behavior confirms the major role played by basic nitrogens that can easily switch from HB acceptors to HB donors depending on the protonation state. In the preliminary work a conspicuous number of outliers having the flavonoid and steroid chemotypes, respectively, have been detected. These structures had been previously characterized as binders to a cytosolic mouse ABCB1 domain vicinal to the ATP binding-site.¹⁰⁸ Moreover, flavonoids also bind to the cytosolic domain of other proteins (e.g. kinases), and it is therefore not surprisingly that they act in a promiscuous manner.¹¹³ Polli et al. previously showed that testosterone non-competitively inhibits Pgp without stimulating ATP hydrolysis.³⁸ For these reasons a separate local model was built, limited to those molecules containing at least one of an ensemble of substructures characteristic for the flavonoid and steroid chemical classes, respectively. We refer to this model as the non-competitive binding model, for lack of a better description. Permeability and hydrophobicity appear to be determinant, while the contribution of size and flexibility appeared negligible; these results partially agree with the work of Boccard et al.¹¹⁴

The composite model has high accuracy with a good balance between specificity and sensitivity, suggesting that for comparable physicochemical features the pharmacophore description becomes critical, and vice-versa. Overall, the Volsurf+ and FLAP-based models are complementary, and were used in combination to provide a highly predictive composite model.

Conclusion

Drug-drug interactions via Pgp inhibition are of great interest in the pharmaceutical sector, as well as drug discovery. Especially when co-administered with a CYP3A4 substrate, a Pgp inhibitor might alter the bioavailability of such substrates, leading to altered pharmacokinetic and pharmacodynamic profiles, as well as overdosing and toxicity. When integrated with in vitro assays, such in silico tools can be valuable during the early drug discovery phases for the early identification of potential drug-drug interactions and toxicity. We developed a novel approach for predicting Pgp inhibition based on the synergistic combination of specific (pharmacophore) and non-specific (general physico-chemical) descriptors. Given the frequent occurrence of Pgp inhibition in new chemical entities, guidelines to reduce toxicity are of great interest. The model presented here suggests that modulation of permeability, flexibility or hydrophobic surface area, might provide ways to overcome this particular issue. For candidates having optimal ADME properties but inhibiting Pgp, minimal structural changes would be required. For this purpose a pharmacophore model for competitive inhibition was elucidated. Interestingly shape appears to emerge as a crucial factor, indicating the importance of the three-dimensional description for antitarget modeling.

From the point of view of Pgp as a target, rather than antitarget, a potent non-toxic inhibitor able to increase the anti-cancer activity of co-administered drugs has yet to be found. Recurrent failures in clinical trials were often related to side effects, since chemosensitizers were initially derived from scaffolds that are active toward other targets.

Since Pgp has dual role as both target and antitarget, the model presented here has double utility, since it can be used to help prevent unwanted effects when evaluating ADME-Tox properties, as well as when searching for new inhibitors when addressing Pgp as target. The performance observed from the training set and validation sets (Internal and External) is supported by a blind prediction, which was confirmed on a set of experimentally tested molecules, since all predictions were correct. Interestingly, the three molecules predicted as non-inhibitors and four of the six predicted as inhibitors were experimentally confirmed, whereas the remaining two are in the range for which we defined the experimental uncertainty for Pgp inhibitor/non-inhibitor classification. We note that two of these newly reported Pgp inhibitors, Ebastine and Aripiprazole, are in the micromolar range.

In conlusion, the uniqueness of the model presented here is the way in which the problem is handled. In fact, the well-known complexity of Pgp inhibition has been split in several issues, which include: cell permeability, pharmacophore identification (for direct binding) and non-competitive binding (probably in the ATP binding site). In this paper we have presented how all these features/issues should be modelled using a sequence of filters and local models and not with a unique model.

Experimental Section

1. Molecules

The molecules of the blind set were purchased from Sigma-Aldrich,¹¹⁵ with the exception of Cyclosporine, which was purchased from Alexis Biochemicals (San Diego, CA, USA) and Aripiprazole, which was obtained from Sanofi Aventis, Bridgewater, NJ, USA.

Active compounds (Aripiprazole, Ebastine, Sertindole, Repaglinide, Ziprasidone, and Aprindine) have been characterised by mass spectrometry; purity was found to be over 95% for all compounds tested.; the corresponding analytical data have been reported in table S4 of Supporting Information.

2. Cells line and cultures

The L5178Y mouse T-lymphoma cell line transfected with a recombinant MDR1/A retroviral vector (pHa MDR1/A), a generous gift from Dr. Michael M. Gottesman (National Cancer Institute, Bethesda, MD, USA), were used. Human MDR1-expressing cells were selected by culturing the transfected cells with 60 ng/ml colchicine to maintain the expression of the MDR phenotype.¹¹⁶ The L5178 MDR1 cell line was grown in McCoy's 5A medium supplemented with 10% heat-inactivated horse serum, 2 mM L-glutamine, 100 UI/ml penicillin and 0.01 mg/ml streptomycin. Cells were maintained in a humidified incubator with an atmosphere of 95% air and 5% CO₂ at 37 °C. When the cells reached confluency, they were harvested and plated for subsequent passages (up to 20) and for drug treatment. Cultures were initiated at a density of 2×10⁵ cells/ml and grown exponentially to about 2×10⁶ cells/ml in 48h. Cells will werebe counted in a Burker cytometer before use and their viability tested by Trypan Blue exclusion.

3. Cell loading with Rhodamine 123 and inhibition of Pgp-mediated Rhodamine 123 efflux assay

The capability of the compounds to inhibit Rhodamine 123 (R123) efflux was determined as already described by Saponara et al.¹¹⁷ with some modifications. Briefly, L5178 MDR1 cells (2×10⁶ ml⁻¹) were resuspended in serum-free McCoy's 5A medium, and 0.5 ml aliquots of the cell suspension were distributed into Eppendorf centrifuge tubes. Compounds tested were added at different concentrations and samples were incubated for 10 min at room temperature. Then, R123 was added to the samples at a final concentration of 5×10⁻⁶ M R123 and cells were incubated for 20 min at 37 °C. Thereafter, cells were washed twice by centrifugation for 5 min at 2,000 g and resuspended in 0.5 ml phosphate-buffered saline (PBS). R123 retained by cells was quantified as fluorescence, using a Becton-Dickinson FACS Calibur flow cytometer (San Josè, CA, USA) equipped with an ultraviolet argon laser (excitation at 488 nm, emission at 530/30 and 585/42 nm band-pass filters). FACS analysis were gated to include only individual, viable cells on the basis of forward and side light-scatter and were based on acquisition of data from 10,000 cells. Fluorescence signals were analyzed by the BDIS CellQuest software (Becton Dickinson, San Josè, CA, USA). The mean fluorescence intensity (MFI) were used for comparison among different conditions. Sodium orthovanadate (Vi) were was selected as the positive control for a standard inhibitor since already at 5×10⁻³ M concentration it can maximally inactivate the Pgp efflux pump.¹¹⁸ IC₅₀ for Vi determined on L5178 MDR1 cells was about 7×10⁻⁴ M.

4. Data analysis and statistics

Data were reported as mean ± SEM of at least three independent experiments run in triplicate. The fluorescence data were expressed as the mean of arbitrary fluorescence units derived from histogram plots of the 10,000 cells that are examined. The percent Pgp inhibition exerted by a single compound was calculated as described by Wang et al.¹¹⁸ The relative fluorescence (i.e. percent inhibition of Pgp) was calculated as mean fluorescence intensity (MFI) of a discrete sample divided by the MFI in the presence of 5×10⁻³ M Vi, times 100:

Relative fluorescence = [MFI of sample/ (MFI of sample + Vi)] × 100

The denominator represents MFI of the sample when inactivation or complete preclusion of the function of Pgp active efflux is attained. The numerator is the resulting signal caused by test compound inhibiting the function of Pgp active efflux. Pgp blocking activity was described by α_max, which expresses the efficacy of the inhibitor and by IC₅₀, which measures its potency. α_max varies between 0 (in the absence of the inhibitor) and 1 (when the amount of R123 found in L5178 MDR1 cells is comparable to that determined in the presence of 5×10⁻³ M Vi). IC₅₀ measures the potency of the inhibitor and represents the concentration that causes a half-maximal increase (α=0.5) in intracellular concentration of R123. IC₅₀ values were obtained by best fitting the concentration-dependent inhibition curves (GraphPadPrism5 program, GraphPad Inc., USA).

Figure 2.

Flow chart of the GRID-based technologies used. While VolSurf+ elaborates MIFs to compute holistic molecular descriptors suitable for modeling pharmacokinetic properties, FLAP extracts hotspots from the GRID MIFs, combining them in quadruplets. Each quadruplet of the molecule A is overlaid with each quadruplet of the molecule B and for each match a score associated to the 3D-superposition of the two molecules is calculated which estimates how much the MIFs are intersected. At the end of the process, only the best score obtained, which corresponds to the best superimposition, is used to assess the 3D-similarity between the two molecules.

Figure 5.

Concentration-response curve of the inhibition of Pgp-mediated R123 efflux from L5178 MDR1 cells by Sertindole (panel A), Ebastine (panel B) and Aripiprazole (panel C) IC₅₀ for the Pgp inhibitors was reported in Table 3. Data are depicted as mean ±e.s.m. of at least three independent experiments run in triplicate.

Table 3.

Inhibition of Pgp-mediated cell efflux of R123 in mouse T lymphoma L5178 MDR1 cells by the nine molecules of the blind set, Cyclosporine A used as control.

Moleculea	Prediction (Classification)	Exp data
		αmaxb	IC50 (M)c	Classificationd
Cyclosporine Ae	N.A.e	0.92f	6.5×10−7	Pgp Inhibitor
Aripiprazole	Pgp Inhibitor	0.99f	7.8×10−7	Pgp Inhibitor
Ebastine	Pgp Inhibitor	0.99f	1.5×10−6	Pgp Inhibitor
Sertindole	Pgp Inhibitor	0.74f,g	6.5×10−6f	Pgp Inhibitor
Repaglinide	Pgp Inhibitor	0.35	N.D.h	Pgp Inhibitor
Ziprasidone	Pgp Inhibitor	0.19	N.D.h	Pgp weak inhibitor
Aprindine	Pgp Inhibitor	0.16	N.D.h	Pgp weak inhibitor
Alprenolol	Pgp Non-Inhibitor	0.05i	N.D.h	Pgp Non-Inhibitor
Nateglinide	Pgp Non-Inhibitor	0	N.D.h	Pgp Non-Inhibitor
Ticarcillin	Pgp Non-Inhibitor	0	N.D.h	Pgp Non-Inhibitor

^aMolecular structures are given as Supporting Information.

^bα_max, expresses the efficacy of the inhibitor and it varies between 0 (in the absence of the inhibitor) and 1 (when the amount of R123 found in L5178 MDR1 cells was comparable to that determined in the presence of 5×10⁻³ M sodium orthovanadate). The values reported represent the efficacy determined at 1×10⁻⁴ M inhibitor concentration.

^cIC₅₀ measures inhibitor potency and represents the concentration that causes a half-maximal increase (α=0.5) in intracellular concentration of R123.

^dThe classification was coherent with the thresholds used in the dataset collection.

^eCyclosporine was used as control. Preliminary data were presented at XIV National Congress of PhD Student in Pharmacological Sciences, Siena (Italy), 20–23 September 2010. It was in the Training set, so no prediction is provided.

^fSee also Figure 5.

^gapparent IC₅₀.

^hN.D. = Not Detectable.

Abbreviations

ADRs

adverse drug reactions

Pgp

P-glycoprotein, also known as ABCB1

ABC

ATP-binding cassette

MDR

multiple drug resistance

hERG

human ether-a-go-go-related gene

CYP3A4

cytochrome P450 3A4

HB

hydrogen bond

MIF

molecular interaction field

FLAP

fingerprints for ligands and proteins

MDC

most descriptive compounds

PCA

principal component analysis

HB

hydrogen bond

QSAR

quantitative structure-activity relationship

ADME

absorption, distribution, metabolism, excretion

PLS-D

partial least square discriminant

LV

latent variable

LDA

linear discriminant analysis