Abstract
Recently, two new classes of reversible inhibitors, the benzbromarones (BZBRs) and the N-3 substituted phenobarbitals (PBs), were used to study the active site characteristics of CYP2C9 and 2C19, respectively. Since these ligands are some of the first CYP2C ligands to extend into the low nanomolar Ki range (Ki < 100 nM), they were subjected to three-dimensional quantitative structure-activity relationship (3D-QSAR) analysis. Given that BZBRs or the PB ligands bind very tightly, it can be assumed that these structures complement the binding pocket(s) for these enzymes. Thus, the resulting models should output a 3D arrangement of interaction sites predicted to be important for near optimal binding to the CYP2C9 and CYP2C19 enzymes. These predicted interaction regions may then improve the ability to predict drug-drug interactions. The resulting models generated from these new high affinity ligands are discussed, as are novel uses of 3D-QSAR and molecular modeling techniques that may be useful in the study of cytochromes P450 specifically.
In ligand-based molecular modeling, knowledge accumulated from experimentation is used to build and test potential models for predicting ligand-protein interactions and hence drug-drug interactions, the sites of drug metabolism, toxicity, and other parameters. Recently, CYP2C9 and CYP2C19 inhibitors with 1 to 2 orders of magnitude lower Ki values than previously characterized compounds have been reported (Suzuki et al., 2002; Locuson et al., 2003). These new inhibitors should help expand various models for CYP2C metabolism. Several groups have published 3D quantitative structure-activity relationship analysis (3D-QSAR) models for P450s, including multiple models for CYP2C9. Therefore, it is relevant to address how the newer, more potent inhibitors, which fit the same alignment rules used previously, have altered existing QSAR models for CYP2C9 and, in some cases, provided distinctions between the chemical properties that appear to define inhibitors of CYP2C9 and CYP2C19. Finally, new applications of 3D-QSAR modeling and how it may enhance P450 metabolism research is discussed.
What Can High Affinity Ligands of P450s Tell Us?
The pursuit for, and study of, high affinity P450 ligands remains an important endeavor. High affinity ligands for each P450 enzyme can help define the enzyme in the form of a pharmacophore, improving the identification of drug leads with the highest potential for drug-drug interactions based on their structure. In practice, this is a lofty goal because determining drug interaction potential in any quantitative manner requires an accurate, universal binding model that can predict any compound's affinity for a given enzyme. Establishing such a model, however, can be difficult because: 1) models used for prediction can only be based on what is currently known, and 2) P450s display complex binding behavior and conformational flexibility. Investigators may have the ability to screen new drugs for their drug interaction potential with the use of selective substrates or inhibitors; however, this is only a screen and not a de novo prediction method allowing elimination of compounds before spending time and resources on synthesis and in vitro screening. It is here that the utility of in silico models can be realized, but the first part of the problem is that models can only be built from currently available information. In silico predictions are based on a postulated enzyme-substrate interaction(s) derived from the target protein's structure, or physicochemical properties of a series of ligands that correlate to some experimental finding (e.g., measure of binding energy or inhibition). One way to improve the predictive ability of models is to develop higher affinity P450 ligands that better describe an enzyme's ideal pharmacophore. Of course, the second issue affecting in silico predictions must be remembered: P450 behavior is often complex. For instance, P450 ligands can retain rotational freedom inside an enzyme active site, as evidenced by the production of multiple oxidation products, and can complex with the heme iron or form adducts with the heme or protein as a result of catalysis in a substrate-dependent manner (reviewed in Meunier et al., 2004). This interesting chemistry is not necessarily related to a pharmacophore based on affinity for protein (but it does represent an important, growing field).
Benzbromarone (BZBR), a uricosuric agent, is a high affinity P450 ligand, and experiments with this agent demonstrate the types of information that can be gained from use of high affinity ligands in P450 research (Fig. 1). First of all, BZBR demonstrates the importance of predicting drug interactions. Shimodaira et al. (1996) reported an interaction of BZBR with CYP2C9 substrate warfarin in vivo. Later, BZBR was confirmed to be a potent competitive inhibitor of (S)-warfarin turnover in vitro by CYP2C9 (Ki < 10 nM), providing evidence that the interaction occurred at the level of the CYP2C9 enzyme (Takahashi et al., 1999). Second, because BZBR was found to be a potent CYP2C9 inhibitor, it could be used to revisit the description of the CYP2C9 pharmacophore. Initially the high affinity of BZBR was puzzling because the most noted property of CYP2C9 substrates is their anionic character. After confirming a submicromolar Ki (19 nM) for BZBR versus (S)-warfarin in a reconstituted system with purified enzymes, a suggestion by an astute advisor led to a pKa determination by 13C NMR and UV absorbance that demonstrated that BZBR's phenol was acidic (Locuson et al., 2003). Hence, the previously suggested hypothesis of a complementary cationic site existing within CYP2C9 did not have to be discarded. More importantly, a new feature was found that had a significant effect on affinity. The subsequent synthesis of BZBR analogs demonstrated that BZBR's affinity for CYP2C9 was substantially derived from halogen substitution, pointing out a specific, previously unrecognized, hydrophobic enzyme-ligand interaction (Locuson et al., 2004a). Moreover, all experimental evidence obtained thus far has shown BZBR to be a reversible and competitive inhibitor that does not derive affinity through heme ligation (Dickmann et al., 2004).
Through use of an interesting set of BZBR derivatives, it was then possible to move from a set of qualitative enzyme-ligand interactions into a quantitative 3D-QSAR model. Various predictive software that carry out 3D-QSAR or molecular dynamics methods to compute binding energies are available to model ligand affinity. With respect to modeling P450s, the reader is referred to a review published in this journal by Ekins et al. (2001). In the case for BZBR, the 3D-QSAR method CoMFA was used and compared with the first CYP2C9 3D-QSARs, which were also constructed using CoMFA. CoMFA has the ability to predict affinity based on structure and the known, relative affinities for a series of ligands without any direct modeling of the protein target. This comparison between BZBR and previous CYP2C9 models could be carried out because the same alignment rules were used for all of the CYP2C9 ligands as described in Fig. 2. Basically, it was found that overlaying the site of metabolism of BZBR with those of the previous models allowed good overlap of the remaining structures as shown previously (Rao et al., 2000). A separate binding mode for the warfarin analogs as suggested by the warfarin complex structure (Williams et al., 2003) was not used since current methodologies do not allow determination of the time a ligand remains at distinct binding sites. Although the ionization state of the BZBRs was carefully tested, it was found that evaluating all ligands in their neutral state did not alter the validity of the models.
Statistically, BZBR CoMFA models were equivalent to earlier models (Rao et al., 2000) constructed with data from warfarins, sulfaphenzoles, NSAIDs, and phenobarbitals using the same alignment rules. Table 1 demonstrates that the q2 values, which determine model validity and, in part, predictive ability, did not fluctuate dramatically. BZBRs did extend the range of included Ki values nearly 2 orders of magnitude on the low end. When the BZBR data were added to previous models containing a wide array of CYP2C9 ligands, the resulting qualitative pharmacophore appeared to improve (Fig. 2). The steric interaction common to both the BZBR model and the previous model for CYP2C9 ligands shown in yellow (Rao et al., 2000) was predicted in the combined model. In addition, one steric and one electrostatic interaction from the earlier model and two hydrophobic interaction sites from the BZBR model were predicted in the combined model. This demonstrated that even though the inhibition constant of BZBR is at least 1 order of magnitude lower than the most potent non-BZBR inhibitors, the combined CYP2C9 pharmacophore represented not just the high affinity BZBR ligands, but the lower affinity ligands as well. Contributions of BZBRs to the combined model, though, were quite striking when represented as a fraction of the model. The two distinctive favorable hydrophobic interactions at the two bromine atoms (42% of total contribution) accounted for as much as the two electrostatic sites (41%), reflecting an improved model. Not only did the model predict the importance of an electronegative group common to CYP2C9 ligands (without requiring knowledge of the ionization state), but it also suggested that the new, heavily weighted hydrophobic interactions would now allow the prediction of unexpectedly low Ki ligands with higher pKa values, such as the dimethyl-substituted BZBR analog (Fig. 1). It should be noted that the ligand alignment rules used for this particular model and outlined in Fig. 2 represent just one possibility; however, the superimposition of the site of metabolism remains a very simple rule and is a property that is usually known early on in drug discovery.
Parallel SAR methodology was being applied to CYP2C9's closest human relative, CYP2C19, by Rettie and coworkers, who modified hydantoins and barbiturates to develop low nanomolar CYP2C19 inhibitors similar in structure to the model substrate mephenytoin (Suzuki et al., 2002). These studies provided an excellent opportunity for comparing and contrasting two related enzymes because inhibition data were collected for the barbiturates against CYP2C9 and for BZBRs against CYP2C19. As stated above, comparison of the models for the two enzymes was possible because there was insufficient evidence to suggest that they required different ligand alignments and because of the enzymes' high sequence identity. In addition, the primary site of metabolism of BZBR by CYP2C19 was the same as that of CYP2C9. For CYP2C19, the alignment rules for the N-3 substituted PB analogs again superimposed the sites of metabolism and barbiturate rings so that the N-3 substituent overlapped in most cases. The site on the molecule where CYP2C9 metabolizes these compounds is unknown. If the site of metabolism were the N-3 substituent instead, then following the same alignment rules would not change the results of the model (i.e., it would be equivalent to rotating all the ligands equally in space, which does not affect CoMFA). The resulting CoMFA models demonstrated distinct differences in the ability of QSAR descriptors to define the data set (Suzuki et al., 2004). As with BZBR/2C9, the high affinity N-3 substituted phenobarbital (PB) series affirmed earlier observations regarding the CYP2C19 pharmacophore and, as well, highlighted the important role of ligand hydrophobicity in affinity. Namely, CYP2C19 demonstrated greater stereoselectivity of the PB's C-5 position than was observed with CYP2C9. Differences in the effects of N-3 group substitution were also observed and were attributed to their 3D hydrophobicity. Conversely, the CYP2C9 model better predicted its PB-derived Ki values based on nonspecific hydrophobicity (i.e., log P value), which is not a 3D descriptor. This is the opposite of the BZBR/2C9 results, where the specific placement of hydrophobic groups was a better descriptor than log P; however, close structural analogs are more adequately described by simple hydrophobicity descriptors, and it is unlikely that such simple descriptors can be used for more diverse data sets. In summary, the PB CoMFA models for CYP2C9 and CYP2C19 echo the importance of hydrophobic substituents in inhibitor potency for CYP2C enzymes, and are attractive because they span a wide range of Ki values and account for CYP2C19's stereochemical selectivity.
CoMFA models for CYP2C19 based on BZBR analogs only give statistically valid results when PBs are included. Entering the BZBR analogs with their correct charge state, which is shown to be more important for CYP2C19 than for CYP2C9, still did not allow the derivation of any valid models. However, when the PBs are included, four distinct ligand alignment schemes all provided statistically valid models, which supports the fact that both classes of ligands are metabolized at both aromatic rings, even though there is always a preference for one site over another. Predicted electrostatic and steric interaction sites, however, were nearly indistinguishable from each other, so that the qualitative pharmacophore aspect of these models, except for stereochemical preference, remains in question.
It appears that 3D-QSAR modeling of higher affinity ligands of P450s may improve our ability to predict drug-drug interactions in several ways. For CYP2C9 and CYP2C19, the inclusion of higher affinity ligands did improve 3D-QSAR models quantitatively by improving our ability to describe the binding of compounds below the micromolar Ki range, an important improvement. Presumably, this includes the ability to rank the effects of stereochemistry and hydrophobicity (3D) or lipophilicity (log P) on the affinity of compounds similar to the BZBRs and PBs with respect to both enzymes. As for a qualitative picture of enzyme-substrate interactions, the BZBR inhibitors both correlated with previous observations while defining new hydrophobic interactions with CYP2C9 through use of multiple classes of ligands (Fig. 2). For clarification, the described 3D-QSAR models were built with one set of rules for each enzyme, and no claim is made that the models account for the observations of all ligands. In fact, given the potential for formation of multiple metabolites, several control alignments may be needed to interpret one of possibly multiple valid models. Even so, such models could still be used for predictive purposes as long as the alignment rules remain consistent.
Like any method, CoMFA does have drawbacks. Because of the first prediction problem, the results are dependent on the training set, or ligands used to derive the model. The method requires inclusion of several ligands that span a large range of Ki values, as well as strict ligand alignment rules that may be difficult to define without additional metabolism, docking, and/or enzyme mutagenesis data to help render hypotheses. However, as more P450 structures become available and are used more to generate ligand alignments, CoMFA models will improve in their ability to predict ligand affinity. Whereas docking methods can suggest residues that may be important in binding of a single compound, QSAR methods like CoMFA are capable of distinguishing the specific residues that are most important for binding given a series of ligands. Thus, rather than move away from the use of QSAR analyses like CoMFA as more P450 enzyme structures are determined, it may prove beneficial to utilize the methodology even more.
New Uses of 3D-QSAR
It should be realized that QSAR analysis may be used for much more than relative affinity prediction in the P450 field, since more may be occurring than just ligands simply moving in and out of these enzymes. For example, we have recently tried to derive QSAR models for the multiple branch points in the P450 catalytic cycle that could lead to uncoupling. Preliminary results suggest promise in modeling of H2O2 formation by CYP2C9. First, the formation of H2O2 in the presence of 22 diverse CYP2C9 substrates was determined using a xylenol orange assay in a reconstituted in vitro enzyme system. Next, it was discovered that a revised ligand alignment based on new crystal structure information was required to develop validated CoMFA models. Common electronegative groups and nearby hydrophobic groups were aligned with highest priority according to the features of the new CYP2C9-flurbiprofen complex structure (Fig. 3) (Wester et al., 2004). No docking was used, but it was predicted that polar groups would interact with R108 and hydrophobic groups would interact with F114 or F476. This was followed by the very loose positioning of the site of metabolism. In other words, little modification of the AM1 calculated geometry was made, even if the primary alignment rules did not allow the sites of metabolism to overlie each other. One notable feature of the predicted interaction sites is that one of these sites (shown in green, Fig. 3) is near the site of metabolism, suggesting that the position of certain steric features of substrates in relation to the heme affects the level of H2O2 formation. A second feature of the model is a favorable electropositive interaction shown in blue near the ligands' electronegative groups (Fig. 3), which suggests that regions further removed from the heme may affect the coupling of catalysis. These previously unpublished CoMFA models possessed cross-validated q2 values between 0.50 and 0.65 and predicted both steric (30%) and electrostatic features (70%). Figure 4 is a plot of one of these cross-validated models. Although it used diverse chemical structures, the project was made difficult by the limited range of H2O2 formation rates (range of ∼30 nmol/min/nmol P450). Only when H2O2 formation as a fraction of NADPH consumed was used as a parameter could models be generated; however, this may have turned out to be a way to add the additional dimension of depth that was needed.
As mentioned previously, CoMFA is only one of several software programs used for 3D-QSAR analysis, but those interested can read Sutherland et al. (2004), which compares CoMFA with closely related programs. Other possibilities for modeling studies include the move toward methods that do not require detailed ligand alignments (Balakin et al., 2004) and incorporate more information derived from determined protein structures and/or homology models. Improved homology models are also leading to more progress in molecular dynamics approaches in describing human P450 isoforms whose structure has not yet been determined (e.g., 2D6) (Venhorst et al., 2003; Kemp et al., 2004). One intriguing approach even uses 3D-QSAR to refine homology models before they are used for further docking to improve the ligand alignments (Evers et al., 2003).
Conclusions
Today a cursory browsing of the literature results in many examples of in silico methods used in drug metabolism or design and their constant evolution. Fortunately, some groups involved in creating these ligand docking and QSAR programs are making them easier and faster to use so that they are more approachable by those less familiar with modeling. Having said that, the roles of P450 motion, water molecules in the active site, and protein-protein interactions are likely tied to their activity in a substrate-dependent manner, requiring much more extensive simulations coupled with detailed biochemical studies. Nonetheless, 3D-QSAR analysis should remain a useful method in P450 studies given these enzymes' unique properties. Certainly, as higher affinity ligands are discovered for each of the P450 isoforms, more will be learned about enzyme specificity, providing a starting point for more extensive analyses and improved predictive capabilities. There is also further room to evaluate the use of 3D-QSAR in defining distinct ligand binding sites for heteroactivators that are occupied in the presence of certain substrates simultaneously (Egnell et al., 2003). Furthermore, a 3D-QSAR treatment for compounds that coordinate to the heme iron will be required to create models that can distinguish between affinity derived from protein contacts and that from heme coordination. In fact, any other P450 behavior that is ligand-dependent, such as the shunting of electrons to H2O2 formation, should lead us to further intriguing results.
Acknowledgments
This review is dedicated to my thesis advisor Dr. Jeff Jones, who challenged me, provided all the resources for my research projects, and critiqued the manuscript. I thank Dr. Tim Tracy for assisting with the proofreading of the manuscript.
Footnotes
-
This work was supported by National Institutes of Health Research Grants ES09122 and GM061823.
-
Article, publication date, and citation information can be found at http://dmd.aspetjournals.org.
-
doi:10.1124/dmd.105.004325.
-
ABBREVIATIONS: 3D-QSAR, three-dimensional quantitative structure-activity relationship; BZBR, benzbromarone; PB, N-3 substituted phenobarbital/nirvanol.
- Received February 17, 2005.
- Accepted April 14, 2005.
- The American Society for Pharmacology and Experimental Therapeutics