De Novo Prediction of P-Glycoprotein-Mediated Efflux Liability for Druglike Compounds

Abstract

P-glycoprotein (Pgp) is capable of recognizing and transporting a wide range of chemically diverse compounds in vivo. Overcoming Pgp-mediated efflux can represent a significant challenge when penetration into the central nervous system is required or within the context of developing anticancer therapies. While numerous in silico models have been developed to predict Pgp-mediated efflux, these models rely on training sets and are best suited to make interpolations. Therefore, it is desirable to develop ab initio models that can be used to predict efflux liabilities. Herein, we present a de novo method that can be used to predict Pgp-mediated efflux potential for druglike compounds. A model, which correlates the computed solvation free energy differences obtained in water and chloroform with Pgp-mediated efflux (in logarithmic scale), was successful in predicting Pgp efflux ratios for a wide range of chemically diverse compounds with a R² and root-mean-square error of 0.65 and 0.29, respectively.

P-glycoprotein (Pgp) plays an important role in the in vivo disposition of a wide range of clinically relevant drugs.¹ The impact of Pgp on permeability across the blood–brain barrier (BBB) has received considerable attention within the drug discovery community.² Active efflux by Pgp at the BBB often represents a significant challenge in the discovery and development of agents that require central nervous system (CNS) penetration for target engagement.^3,4 In addition, overexpression of Pgp in cancer tumor cells has been implicated in the development of drug resistance to anticancer agents.⁵

Pgp is a promiscuous transporter that is capable of recognizing and transporting a wide range of chemically diverse structures.⁶ An X-ray crystal structure of mouse Pgp, which shares 87% sequence identity with human Pgp, was recently reported.¹ This structure can provide insight toward understanding the mechanism by which the human Pgp transporter functions. Druglike compounds were hypothesized to be recruited to the Pgp binding pocket while they are either diffusing through the membrane or through the intracellular-facing opening.¹ Upon binding, the intracellular-facing state of Pgp undergoes an ATP-dependent state change that renders it extracellular facing and allows the substrates to be exported into the extracellular domain (Figure 1).¹

Figure 1.

Proposed mechanism for the efflux of compounds by Pgp (adapted from ref (1)). Neutral compounds were assumed to bind to the intracellular-facing state. Upon binding, Pgp undergoes an ATP-dependent state change and releases the compound into the extracellular domain. We hypothesized that the probability of transferring compounds from the Pgp binding pocket to the extracellular domain may be used to predict the Pgp-mediated efflux liability.

Herein, we report our findings, which correlate the susceptibility of a compound to Pgp-mediated efflux with the free energy required for a compound to transition from a lipophilic to an aqueous environment (Figure 2) since the residues surrounding the Pgp drug binding pocket are mainly hydrophobic and aromatic. We proposed that a significant element contributing to efflux is a compound's preference to be in an aqueous-like environment and (heuristically) hypothesized that Pgp-mediated efflux may be related to the probability by which a compound will be transferred from Pgp binding pocket to the extracellular domain.⁷ Accordingly, Pgp transports druglike compounds to the extracellular domain when there is a strong preference for the compound to be in an aqueous-like environment. Alternatively, when the preference to be in the extracellular domain is diminished, the ability for Pgp to transport compounds to the extracellular domain is also diminished. This new hypothesis significantly differs from existing hypotheses about how Pgp binds or transports druglike compounds, which rely on recognition of polar functional groups such as amides, hydrogen bond donors, or acceptors; we speculated that these metrics may be seen as surrogates for the hydrophilicity of compounds. However, current hypothesis does not invalidate any of the literature reported hypotheses. Because the extracellular-facing state of Pgp is unknown, we proposed that the relative affinities for the compounds to be in the Pgp binding pocket (i.e., hydrophobic environment) or in the extracellular domain (i.e., aqueous-like environment) may be approximated with solvation free energies computed in chloroform (G_CHCl3) and water (G_H2O), respectively (see the Experimental Procedures for details.) We computed G_CHCl3 and G_H2O for neutral species. In cases where compounds contain ionizable functional groups, such as basic amines or carboxylic acids, we utilized computed pK_a values to estimate the free energy difference between neutral and ionized forms in an aqueous environment and corrected the G_H2O values with the energy difference obtained with the expression ΔG = −1.4(pH – pK_a). This approximation allowed us to test the hypothesis outlined in Figure 1 by using G_CHCl3, G_H2O, and pK_a values computed for druglike compounds.

Figure 2.

Relationship observed between net ER (logarithmic scale) and computed solvation free energy difference, ΔG_H2O–CHCl3. The dashed line is the fit obtained for the data with eq 1. Green, compounds with net ER < 2; red, compounds with net ER ≥ 2; open circles, training set; and closed circles, test set.

Solvation free energy differences between water and chloroform (ΔG_H2O–CHCl3) were computed for a diverse set of 282 compounds and were plotted against the logarithm of net efflux ratio (ER) (Log[net ER]) in Figure 2 (see the Experimental Procedures for details.) Net ER was calculated by dividing the ER obtained in Pgp overexpressing cell lines with the ER obtained in parental cell lines. There was a sigmoidal relationship between Log[net ER] and ΔG_H2O–CHCl3, and this relationship can be described with eq 1, where a = −1.17, b = 0.78, c = −13.90, and d = 1.22 (see the Supporting Information for the relationship observed between Log[ER] and ΔG_H2O–CHCl3). The R² and root-mean-square error obtained from such a fit for training (test) set were 0.65 (0.64) and 0.28 (0.29), respectively. Reduced χ² for this fit for all of the compounds was 0.9.⁸ There were 134 and 148 compounds in the training and test sets, respectively. In this plot, a large negative ΔG_H2O–CHCl3 suggests a strong preference for the compounds to be in an aqueous environment, while a small negative ΔG_H2O–CHCl3 suggests a weaker preference for such media. The relationship observed in Figure 2 allowed us to classify compounds into three categories: (i) compounds that are likely to be Pgp substrates and predicted to be Pgp substrates with high confidence, ΔG_H2O–CHCl3 < 16 kcal/mol (51 compounds); (ii) compounds that are unlikely to be Pgp substrates and predicted not to be Pgp substrates with high confidence, ΔG_H2O–CHCl3 > −12 kcal/mol (75 compounds); and (iii) predictions made with low confidence, −12 < ΔG_H2O–CHCl3 < 16 (156 compounds). The steep relationship observed between Log[net ER] and ΔG_H2O–CHCl3 around −14 kcal/mol limited our ability to make accurate predictions in this range; the method allowed us to accurately predict the efflux liability for 45% of the compounds. It should be noted that the compounds included in this analysis spanned a wide range of ERs and permeability values. While a strong correlation between ΔG_H2O–CHCl3 and Pgp-mediated efflux was observed, no such correlation was observed between ΔG_H2O–CHCl3 and permeability, suggesting that the observed relationship is unique (see the Supporting Information for details).

1

Equation 1 was next used to predict the net ER for 12 clinically relevant drugs (Table 1).⁹⁻¹² The goal of this exercise was not to predict the absolute ER but to test the method's ability to distinguish high efflux compounds from those that would not be expected to succumb to Pgp-mediated efflux.¹³ As seen in Table 1, loratadine, tacrine, and midazolam, drugs reported not to be Pgp substates, were predicted to have low ERs (≤1.2). In contrast, compounds that were reported to be substrates for Pgp were predicted to have higher ERs (1.9–16.7).

Table 1. Predicted and Measured ERs for 12 Drug Molecules.

drug	net ERa	ERpred	GH2O	GCHCl3	pKaACD	ΔGH2O–CHCl3	ref
loratadine	0.5	1.2	–10.43	–2.68		–7.75	(9)
tacrine	1.2	1.1	–12.23	–5.81		–6.42	(11)
midazolam	0.9–1.0	1.2	–8.81	–1.83		–6.98	(9, 11)
zolmitriptan	2.5	15.4	–26.04	–10.64	9.5	–15.40	(11)
prednisone	2.8	7.3	–23.51	–8.57		–14.94	(9)
prazosin	2.9	11.3	–24.36	–8.20		–16.16	(9)
loperamide	3.8	1.9	–9.44	–0.32	9.5	–9.12	(9)
methysergide	4.3	9.7	–21.12	–5.47	7.2	–15.65	(11)
labetalol	2.4–6.5	12.5	–20.83	–6.91	9.3	–13.92	(10)
quinidine	7.4	5.3	–15.71	–4.52	9.6	–11.19	(9)
ritonavir	18.4	16.7	–23.68	2.17		–25.85	(9)
eletriptan	44.7	11.1	–19.09	–7.18	10.4	–11.91	(11)

^aNet ER was calculated by dividing the ER obtained in Pgp overexpressing cells with the ER obtained in parental cells.

Predictions made for labetalol, midazolam, and zolmitriptan are exemplified in Figure 3. Computed G_CHCl3 and G_H2O for labetalol are −6.91 and −20.83 kcal/mol, respectively; this corresponds to ΔG_H2O–CHCl3 of −13.92 kcal/mol. Labetalol has a basic amine group, and the ACD predicted pK_a for this amine group is 9.3. Because G_H2O computed for the neutral drug underestimates the polarity of labetalol in water, a pK_a-based correction was applied to account for the free energy difference between the ionized and the netural forms; we obtained a pK_a-corrected solvation free energy difference of −13.92 + −1.4 × (9.3 – 7.4) = −16.58 kcal/mol. Plugging −16.58 for ΔG_H2O–CHCl3 into eq 1 gave a predicted ER (ER_pred) of 12.5. Computed G_CHCl3 and G_H2O for midazolam are −6.91 and −20.83 kcal/mol, respectively, and gave a ΔG_H2O–CHCl3 of −6.98 kcal/mol and an ER_pred of 1.2. The method overestimated the ER measured for zolmitriptan; we observed that conformationally flexible molecules with distributed polarity were sometimes predicted incorrectly. The use of conformational ensembles in calculations can potentially alleviate this problem.

Figure 3.

Predicted and measured ERs for labetalol, midazolam, and zolmitriptan. The in-house measured ER for labetalol is 12, and literature-reported values range from 2.9 to 6.0.

A significant amount of work has been directed toward understanding the physicochemical properties or molecular attributes that are associated with Pgp-mediated efflux of druglike compounds. Several physicochemical properties have been reported to correlate with Pgp recognition, including cLog P, topological polar surface area (TPSA), and the number of hydrogen bond donors (HBD) and acceptors (HBA).^14,15 Net ER data for compounds that are likely Pgp substrates (as defined by net ER ≥ 2) and non-Pgp substrates (as defined by net ER < 2) were plotted against cLog P, TPSA, HBD, and HBA for the same set of 282 Amgen compounds (Figure 4). Histograms given in Figure 4A corroborate the observations reported in the literature^14,15 and can provide empirical guidelines to help minimize Pgp-mediated efflux. For example, the following attributes are commonly used to help design low efflux compounds: the inclusion of <2 HBD, TPSA of <70 Ų, and cLog P of >3. While these guidelines can be useful in the drug design process, it is well-known that compounds with the same physicochemical properties can have substantially different levels of Pgp-mediated efflux. Furthermore, these relationships fail to give clear guidelines to medicinal chemists as there are considerable overlaps between compounds with net ER ≥ 2 and net ER < 2 across each of the previously mentioned properties (Figure 4). Additionally, recognition elements required to interact with certain protein targets often do not align with those that correlate with low efflux, necessitating the need for a more refined approach. In addition to the observed relationships between molecular descriptors and ER, numerous in silico models have been developed to predict the efflux liability for druglike compounds.¹⁶⁻²³ Most of these models rely on existing data sets to predict the ER of new compounds based on their similarity to existing compounds with data. While such models are very useful in late stage projects when interpolations are being made, the predictive power of such models becomes questionable when new chemical space is being explored and there is little or no existing data to “train” these models. Recently, ERs of various compounds have been predicted by using a flexible docking method that makes use of the intracellular-facing state of the mouse Pgp protein.¹⁶ While a correlation between docking score and logarithm of ERs was observed, the relevance between the binding score obtained with the intracellular-facing state and active transport by Pgp remains to be explored.

Figure 4.

Relationships observed between measured Net ER and cLog P, HBD, TPSA, and HBA. Green, compounds with net ER < 2; red, compounds with net ER ≥ 2. Dashed lines in panel A are fits for the observed histograms.

In addition to the guidelines mentioned above, several strategies have also been used to help modulate Pgp-mediated efflux while maintaining biological activity. For example, in cases where a HBD is required to maintain key interactions with the biological target of interest, it has been found that the inclusion of strategically placed HBAs or steric bulk can serve to mask the HBD and in doing so mitigate Pgp-mediated efflux liability.²⁴ Although such modifications are made in an attempt to mask polarity, the TPSAs obtained for both analogues may remain the same and provide a disconnect between what is experimentally observed and the probabilistic relationships deduced from plots like those shown in Figure 4A. Table 2 further illustrates this point with several pairs of molecular fragments. Each of these pairs has the same TPSA values but can be differentiated by their calculated solvation free energies. The pyridone (1)/6-methylpyridone (2) pair represents a case in which the polarity of the fragment is aimed to be manipulated with a steric bulk. The 1,3,5-triazine (3)/1,2,3-triazine (4) pair represents a case in which the polarity is distributed differently across the fragments. Compound 4 may be perceived to be more polar due to the close proximity of the heteroatoms. Finally, N-methylpicolinamide (5)/N-methylisonicotinamide (6) represent a pair of compounds wherein the HBD has the potential to be masked by a hydrogen bond acceptor in 5 but not in 6.

Table 2. Fragments That Highlight the Difference between TPSA and Solvation Free Energies.

The pyrazine (7)/pyrimidine (8) pair is another example that shows that the distribution of polarity can have an impact on ΔG_H2O or ΔG_H2O–CHCl3. An analysis of our internal data²⁵ provided 35 pairs of compounds that contain matched molecular 7/8 pairs with measured ER. Out of these 35 pairs, 26 pyrimidine analogues have higher net ERs than their matched pyrazine analogues, three pairs have the same net ERs, and five pyrazine analogues have higher net ERs (net ER_8<7). However, four out of these five pairs with net ER_8<7 had net ERs that are less than 2 and were considered not to be efflux substrates. We also observed that the pyrimidine analogues, on average, had 2-fold higher net ERs (see the Supporting Information for details).

In summary, we have shown that ΔG_H2O–CHCl3 can be used as a de novo descriptor to predict the likely Pgp-mediated efflux liability for a diverse set of druglike molecules and supports the hypothesis outlined in this work. The method described herein can be used to rank order target compounds or make quantitative predictions upon fitting the ER data to ΔG_H2O–CHCl3. Hence, we believe that it is a useful method for predicting efflux liability for druglike compounds when there is little or no data available to guide the medicinal chemistry efforts with in silico models that rely on training sets. The mechanism by which Pgp recognizes or binds druglike molecules is beyond the scope of this work, and the correlation observed between net ER and ΔG_H2O–CHCl3 was not aimed to address this issue.

Experimental Procedures

Passive permeability was determined in LLCPK-1 cell monolayers, while ER (B → A/A → B) was measured in LLCPK-1 cell monolayers that were transfected with human MDR1. All of the compounds included in this analysis demonstrated passive permeability [(A → B + B → A)/2] of >27 nm/s and efflux of <2.8 in the nontransfected cell line. These measures ensure that the efflux observed in the LLCPK-1 cells transfected with MDR is a robust measurement of efflux mediated by Pgp. Compounds included in this analysis were selected from six different programs and include 10 different subseries. See the Supporting Information for tanimoto similarities among all possible molecular pairs. ERs of the compounds in the data set range from 1 to 94.

G_H2O and G_CHCl3 were obtained with single point solvation free energies for the ground state conformations of the compounds obtained from molecular mechanics minimizations with MMFF94x force field with the Born solvation energies. G_H2O and G_CHCl3 were computed with the HF/6-31+G(d) method by using Gaussian 03²⁶ software; this method was shown to produce accurate solvation free energies.²⁷ The calculation of solvation free energies for one conformation introduces errors. To estimate the magnitude of the errors introduced by using only one molecular mechanics conformation, we calculated the Boltzmann weighted solvation free energies [for a conformational ensemble obtained at the B3LYP/6-31G(d) level] for 10 randomly selected compounds. Even though Boltzmann weighted G_H2O and G_CHCl3 were different from those obtained for molecular mechanics obtained ground state conformations, there was a good correlation when Boltzmann weighted ΔG_H2O–CHCl3 was plotted against ΔG_H2O–CHCl3 obtained for molecular mechanics ground state conformation (see the Supporting Information for details.) The use of other organic solvents such as benzene and chlorobenzene produced similar results. We chose to use the ΔG_CHCl3 due to the close proximity of the chloroform dielectric constant to the dielectric constants of the residues surrounding the Pgp binding pocket. The Supporting Information contains detailed description around the diversity metrics of the compounds used in this work and the derivation of the relationship between ER and the computed solvation free energy differences.

Glossary

Abbreviations

Pgp

P-glycoprotein

BBB

blood–brain barrier

CNS

central nervous system

ER

efflux ratio

G_H2O

solvation free energy in water

G_CHCl3

solvation free energy in chloroform

ΔG_H2O–CHCl3

solvation free energy difference between water and chloroform

HBD

hydrogen bond donor

HBA

hydrogen bond acceptor

TPSA

topological polar surface area

Supporting Information Available

Permeability vs ER and cLog P relationships, metrics around the diversity of the compounds, derivation of the relationship between ER and ΔG_H2O–CHCl3 and XYZ coordinates for the drug molecules in Table 1, an example Gaussian input file, and the presentation of the results in the form of confusion plot. This material is available free of charge via the Internet at http://pubs.acs.org.

The authors declare no competing financial interest.