Introduction

Top Abstract Introduction Materials & Methods Results Discussion References

An understanding of the substrate specificity of the cytochrome P-450 (CYP) enzymes is essential due to their clinically important role in the metabolism of xenobiotics and endobiotics (Wrighton and Stevens, 1992). There is no consensus regarding a 3-dimensional (3D) structure for any membrane-bound mammalian CYP active/binding site. Thus, our understanding of the structures of mammalian CYPs are limited to homology models built using bacterial CYPs as a template (Lewis and Lake, 1997) or models built and inferred from known inhibitor binding (Strobl et al., 1993). Many pharmacological studies have resolved receptor active/binding sites using numerous computational 3D-quantitative structure activity relationship (3D-QSAR) techniques (Koehler et al., 1996). These methods utilize relevant conformers of ligands to suggest functional groups, the geometry of structural features, and regions of electrostatic and steric interactions essential for activity or fit to the receptor binding/active site. The overall combination and 3D spatial distribution of physicochemical properties, the functional groups of the ligands, and a measure of binding site properties of an enzyme, such as the K_{m (apparent)} (Nelsestuen and Martinez, 1997), define the pharmacophore. 3D-QSAR methods can also be applied to modeling CYP substrates for the binding/active site(s). Very few CYP binding/active site models have been constructed which utilize enzyme kinetic values for the site in question; these in turn have dealt solely with inhibitory interactions (Strobl et al., 1993; Jones et al., 1996). In contrast, there has been some application of QSAR for oxidation by CYP which was correlated with hydrophobic and steric parameters of substrates (Hansch and Zhang, 1993; Gao and Hansch, 1996). These same authors discussed the value of using K_m and k_cat and reported that the majority of QSAR studies using these parameters resulted in a better than chance correlation (Gao and Hansch, 1996). Therefore, hydrophobic and steric properties were found to be necessary requirements for CYP mediated metabolism. Molecular electron substituent effects in the oxidation of CYPs were found to oppose enzyme-substrate complex formation and were not necessary; however, this may vary for each CYP (Gao and Hansch, 1996).

It has previously been reported that although CYP2B6 represents a small percentage of total human hepatic P-450 (<0.2%) (Shimada et al., 1994), its ability to catalyze the oxidation of xenobiotics may have been underestimated (Ekins et al., 1997, 1998). Many examples of xenobiotics metabolized by CYP2B6 have been identified (Ekins et al., 1998) and in some cases apparent kinetic parameters are also reported. This kinetic information would allow modeling of common features required of substrates for CYP2B6 (and indirectly provide some details of the active site) if the K_{m (apparent)} values encompassed a sufficient range in affinity for the enzyme in question.

As yet there have been no studies describing absolute structural requirements for substrates or inhibitors nor their interactions with the amino acids in the active site of CYP2B6. A homology model of rat CYP2B1 aligned upon CYPBM-3 has been proposed and the amino acids that interact with substrates have been identified. Based on this homology modeling work with CYP2B1, most substrates for this enzyme appear to be involved in a -stacking arrangement with side chains of Phe-181 and/or Phe-263. These amino acids were aligned in the same positions in homology modeled CYP2B6 (Lewis and Lake, 1997), implying their importance in substrate binding. However, caution should be observed in extrapolating the CYP2B1 homology model to CYP2B6 as these CYPs are only 75% similar by protein sequence (Wolf et al., 1988) and, as observed with other CYPs, species differences in metabolic capability are likely to occur. A clear example of this is cocaine N-demethylation which although mediated by rat CYP2B1 and rabbit CYP2B4 (Lewis and Lake, 1997) is not catalyzed by human CYP2B6 (Poet et al., 1996).

In the current study we describe two 3D-QSAR models for CYP2B6 substrates using K_{m (apparent)} values generated with B-lymphoblastoid expressed CYP2B6. These 3D-QSAR models were then used to predict the K_{m (apparent)} values of a test set of five CYP2B6 substrates not present in the training set. The predictions made by both models were then compared with the experimental values of K_{m (apparent)} derived by ourselves or others; this enabled a residual to be calculated. Both models generated in this study may be used in the future to predict substrate affinity of additional molecules. In addition, the pharmacophore model will eventually allow database searching and construction of a receptor site model (Hahn and Rogers, 1995) to confirm features essential for substrate binding in the active site of CYP2B6.

	Materials and Methods

Top Abstract Introduction Materials & Methods Results Discussion References

Materials. NADPH, lidocaine, amitriptyline, and nortriptyline were purchased from Sigma Chemical Co. (St. Louis, MO). Bupropion, verapamil, and norverapamil were obtained from Research Biochemicals International (Natick, MA). Triethylamine was purchased from Aldrich Chemical Co. (Milwaukee, WI). Monoethylglycinexylidide (MEGX) was a kind gift from Dr. Peter Olinga (Groningen University, The Netherlands). Methanol and acetonitrile were purchased from Burdick and Jackson (Muskegan, MI). Microsomes prepared from human B-lymphoblastoid cells expressing vector only and human B-lymphoblastoid cell lines expressing CYP2B6 were purchased from Gentest Corp. (Woburn, MA).

Molecular Modeling. The computational molecular modeling studies were carried out using Silicon Graphics Indigo and Onyx workstations.

Modeling with Catalyst. The 3D structures of substrates were built interactively using either Catalyst version 2.3 or 3.1 (Molecular Simulations, San Diego, CA). The number of conformers generated for each substrate was limited to a maximum of 255 with an energy range of 20 kcal/mol. Ten hypotheses were generated using conformers for the 16 molecules in the training set and the K_{m (apparent)} values (Table 1) after manual selection of the hydrogen bond donor, hydrogen bond acceptor, hydrophobic, and negative ionizable features. After assessing all 10 hypotheses generated, the lowest energy cost hypothesis was considered the best.

Modeling with Molecular Surface-Weighted Holistic Invariant Molecular and Partial Least-Squares. Each molecule was coded in a simplified molecular input line entry system string format (Weininger, 1988). Atomic 3D coordinates and Gasteiger-Huckel charges were generated by CONCORD 3.2.1 (CONCORD user's manual, Tripos Inc., St. Louis, MO). Molecular surface-weighted holistic invariant molecular (MS-WHIM), descriptors were computed using the program EL3DMD (Bravi et al., 1997; Bravi and Wikel, 1998a,b). MS-WHIM descriptors are a set of statistical parameters which contain information about molecular structure in terms of size, shape, symmetry, and distribution of molecular surface point coordinates after weighted centering and principal component analysis. The following weights are applied: unweighted value, positive electrostatic potential, negative electrostatic potential, hydrogen bond acceptor, hydrogen bond donor capacity, and hydrophobicity; these weights yield a total of 72 descriptors [see Bravi et al. (1997) for details on MS-WHIM].

Prediction with the Test Set Molecules. The test set of five molecules (lidocaine, amitriptyline, bupropion, verapamil, and arteether) not included in the initial training set of 16 molecules, were best-fit to the best Catalyst hypothesis in order to predict the K_{m (apparent)} value. The q² value was used to select the best PLS MS-WHIM model and was successively used to predict the K_{m (apparent)} value of test set molecules. PLS also provides for a rapid evaluation of how test set molecules are structurally similar to the training set, hence how predictions can be considered reliable. This method is based on a comparison between the unexplained variance (s_new²) of the molecule to be predicted and the unexplained variance of the training set (s_0²) by means of an approximate Fisher's exact test (Wold et al., 1993). If F is larger than a critical value with (n  a) and (n  a  1)² degrees of freedom (and a is the number of PLS components), the molecule to be predicted is judged to be dissimilar to the training set. Hence, the prediction by the PLS model can be considered unreliable.

Lidocaine N-Deethylation Assay. Initial rate conditions for the formation of MEGX from lidocaine were determined in preliminary studies with microsomes prepared from cell lines expressing CYP2B6. Microsomes (0.4 mg/ml) were preincubated in plastic Eppendorf tubes for 3 min at 37°C with a final concentration of 100 mM sodium phosphate buffer (pH 7.4) and 1 mM NADPH in a 250-µl final volume. Reactions were initiated by addition of lidocaine (0.25-500 µM), terminated after 60 min by adding 25 µl of 70% perchloric acid and then placed on ice. After centrifugation, a 50-µl aliquot of supernatant was analyzed for MEGX by high-performance liquid chromatography (HPLC).

Amitriptyline N-Demethylation Assay. Initial rate conditions for the formation of nortriptyline from amitriptyline were determined in preliminary studies with microsomes prepared from cell lines expressing CYP2B6. Microsomes (0.4 mg/ml) were preincubated as described above in a 250-µl final volume. Reactions were initiated by addition of amitriptyline (5-1000 µM), terminated after 90 min by adding 125 µl of 20% trichloroacetic acid and then placed on ice. After centrifugation, a 50-µl aliquot of supernatant was analyzed for nortriptyline by HPLC.

Bupropion Hydroxylase Assay. Initial rate conditions for the formation of hydroxybupropion were determined in preliminary studies with microsomes prepared from cell lines expressing CYP2B6. Microsomes (0.6 mg/ml) were preincubated as described above in a 250-µl final volume. Reactions were initiated by addition of bupropion (10-1000 µM), terminated after 180 min by adding 125 µl of acetonitrile and then placed on ice. After centrifugation, a 25-µl aliquot of supernatant was analyzed by HPLC.

Verapamil Metabolism. Initial rate conditions for the formation of verapamil metabolites were determined in preliminary studies with microsomes prepared from cell lines expressing CYP2B6. Microsomes (0.4 mg/ml) were preincubated as described above in a 250-µl final volume. Reactions were initiated by addition of verapamil (5-1000 µM), terminated after 180 min by adding 125 µl of acetonitrile, and then placed on ice. After centrifugation, a 5-µl aliquot of supernatant was analyzed by HPLC.

Kinetic Evaluation and Interpretation. Kinetic analyses of verapamil O-demethylation, amitriptyline N-demethylation, bupropion hydroxylation, and lidocaine N-deethylation were initially assessed by visual examination of Eadie-Hofstee plots to determine whether Michaelis-Menten enzyme kinetics were observed or whether an allosteric activation mechanism was apparent (Enzyme Kinetics 1.6; Window Chem Software Inc., Fairfield, CA). The estimates for the kinetic parameters from these analyses were used as initial estimates for nonlinear regression analyses (NONLIN, version VO2-G-VAX; Statistical Consultants, Inc., Lexington, KY) fitting the data to Michaelis-Menten or Hill equations and generating K_{m (apparent)}, V_{max (apparent)}, and, when appropriate, n (number of sites bound by activator)(Segel, 1993). The best fit to a particular model was determined by examination of (in order of importance) the randomness of the residuals, the size of the residuals, the standard error of the parameter estimates (first between models, then within a model using the best fit weighted model), and finally, when necessary, the difference between sums of squares using Fisher's exact test (Boxenbaum et al., 1974).

	Results

Top Abstract Introduction Materials & Methods Results Discussion References

Catalyst CYP2B6 Substrate Pharmacophore. Catalyst is a commercially available program that enables pharmacophore construction by using a collection of molecules with affinities over a number of orders of magnitude for the enzyme binding site. The pharmacophore hypotheses produced by Catalyst explain the variability of the affinity of the substrates for the active site with respect to the geometric localization of features present in the molecules used to build it.

View larger version (136K): [in this window] [in a new window]   Fig. 1.   The Catalyst-produced CYP2B6 substrate pharmacophore (See Materials and Methods) illustrating three hydrophobic areas (cyan) and a hydrogen bond acceptor feature (green) with a vector in the direction of the putative hydrogen bond. The inter bond angles and the distances between each hydrophobe and the hydrogen bond acceptor are also annotated.

View larger version (156K): [in this window] [in a new window]   Fig. 2.   The Catalyst-produced CYP2B6 substrate pharmacophore with 7-ethoxy-4-trifluoromethylcoumarin (see Materials and Methods) illustrating three hydrophobic areas (cyan) and a hydrogen bond acceptor feature (green).

View larger version (137K): [in this window] [in a new window]   Fig. 3.   The Catalyst-produced CYP2B6 substrate pharmacophore with S-mephenytoin (see Materials and Methods) illustrating three hydrophobic areas (cyan) and a hydrogen bond acceptor feature (green).

View larger version (16K): [in this window] [in a new window]   Fig. 4.   Correlation of Catalyst-produced estimated Km (apparent) values with the observed Km (apparent) values generated with CYP2B6 (see Materials and Methods). The central line corresponds to the regression for the data, the dashed lines represent the 95% confidence interval for the regression while the outer lines are the 95% confidence interval for the population.

PLS MS-WHIM CYP2B6 Substrate Model. It is generally accepted that receptors and substrates recognize each other at their molecular surface (Wagener et al., 1995). Therefore, the binding of a ligand depends on the shape of its surface and on the distribution of certain properties (e.g., electrostatic potential) on this surface. MS-WHIM modeling approaches attempt to capture this information and condense it into a brief numerical vector which can be used for 3D-QSAR purposes. This MS-WHIM approach has been successfully applied to study structure-property and structure-activity problems (Bravi et al., 1997; Bravi and Wikel, 1998a,b).

View larger version (16K): [in this window] [in a new window]   Fig. 5.   Correlation of PLS MS-WHIM-produced predicted Km (apparent) values with the observed Km (apparent) values generated with CYP2B6 (see Materials and Methods). The central line corresponds to the regression for the data. The dashed lines represent the 95% confidence interval for the regression, whereas the outer lines represent the 95% confidence interval for the population.

Test Set Predicted K_{m (apparent)} and Observed K_{m (apparent)}. After constructing the Catalyst and PLS MS-WHIM 3D-QSAR models from the training set of substrates, we were able to use them to predict the K_{m (apparent)} values for a test set of molecules not included in the training set. The K_{m (apparent)} values were predicted for a number of CYP2B6-mediated biotransformation pathways including verapamil O-demethylation, bupropion hydroxylation, amitriptyline N-demethylation and lidocaine N-deethylation. These predicted values were then compared with the corresponding values generated using B-lymphoblastoid expressed CYP2B6 (see below). Recently, arteether was identified as substrate for CYP2B6 and kinetic parameters were calculated after using B-lymphoblastoid expressed CYP2B6 (Grace et al., 1998). The K_{m (apparent)} for arteether biotransformation was also predicted using both the Catalyst and PLS MS-WHIM 3D-QSAR models for CYP2B6.

View larger version (133K): [in this window] [in a new window]   Fig. 6.   The Catalyst-produced CYP2B6 substrate pharmacophore with lidocaine (see Materials and Methods) illustrating three hydrophobic areas (cyan) and a hydrogen bond acceptor feature (green).

	Discussion

Top Abstract Introduction Materials & Methods Results Discussion References

It is possible to define the substrate specificity of the active site of an unknown enzyme by building a pharmacophore that utilizes a number of molecules and their affinities for this site. This approach provides complementary information to modeling the substrate binding site for the CYPs, as no crystal structure for a mammalian CYP has been identified. There have been several attempts to build pharmacophores for a CYP by manual overlapping of substrates for CYP2D6 (de Groot et al., 1997 and references therein) and CYP2C9 (Jones et al., 1996). These along with other indirect QSAR modeling techniques have been used to gain some limited understanding regarding the regio- and stereoselective properties and active site geometries of these enzymes. The present study has described two different 3D-QSAR analyses performed on a training set of 16 CYP2B6 substrates. The creation of these models enabled prediction of the K_m(apparent) for a test set of five CYP2B6 substrates, which were excluded from the training set.

These two computational approaches, Catalyst and PLS MS-WHIM, are two quite different modeling approaches. Within Catalyst, molecules are described in terms of pharmacophoric features, i.e., hydrogen bond donor or acceptor atoms, hydrophobic moieties, and positively or negatively charged groups. The final result is represented by the 3D spatial disposition of essential features for the activity. The advantages of Catalyst are in the ease of visual interpretation of the results, providing complementary information about the active site, relative estimations of activities of a test set, and the potential use of the pharmacophore for either molecule design or 3D database searches. On the other hand, Catalyst attempts to produce a structure-activity correlation considering only the geometrical disposition of structural features. This is something of an oversimplification, as a number of other physicochemical properties and the 3D geometrical disposition of pharmacophoric features possessed by the molecule may also affect the biological activity.

In comparison to Catalyst, MS-WHIM descriptors are aimed at capturing global 3D chemical information at the molecular surface level. The shape and the electronic properties of a molecule are characterized in 3D space and condensed into a brief numerical vector. A structure-activity relationship is derived by means of PLS and the final model is presented in the form of a linear equation. The advantages of this approach are in the high speed of computation and high degree of automation of MS-WHIM descriptors, of which there are fewer than required for comparative molecular field analysis. Moreover, the MS-WHIM approach provides for a global molecular description with results invariant to molecular rotation and translation. Hence, molecules do not need to be aligned before MS-WHIM computation. This is the main advantage of "global" descriptors over "local" descriptors such as those generated by comparative molecular field analysis, whose results are strictly dependent on how molecules have been aligned and superimposed. On the other hand, the mathematical procedure that underlies the calculation of MS-WHIM descriptors makes their physical interpretation almost impossible. Therefore, a 3D display of the PLS MS-WHIM model is not available. As a consequence, MS-WHIM-based models can be solely used to predict the activity of unknown molecules.

In the present study both 3D-QSAR approaches yielded models that were then subjected to validation procedures. The Catalyst pharmacophore consisted of three hydrophobic features that are likely to interact with a hydrophobic binding domain within the active site of CYP2B6. In addition, this pharmacophore possesses a hydrogen bond acceptor feature with a vector representing the projected point location of the target hydrogen bond donor (Fig. 1). However, Catalyst pharmacophore hypotheses were developed even after using permuted K_{m (apparent)} values. This suggests that the CYP2B6 pharmacophore generated is suboptimal. There have been few published reports utilizing Catalyst (Kaminski et al., 1997) and of these none have commented upon the required size of data sets, the hypothesis cost differential compared to the null hypothesis, or statistical analyses necessary to produce valid pharmacophores. It is therefore difficult to put the Catalyst pharmacophore observations completely into context, although it does provide a starting point for future comparisons.

The best PLS MS-WHIM model was obtained using the following molecular surface properties: unitary value (unweighted), positive electrostatic potential, hydrogen bonding acceptor capacity, and hydrophobicity. Based upon statistical significance, the model produced by PLS MS-WHIM appears to be more robust than the Catalyst pharmacophore. The q² obtained by means of two cross-validation techniques, LOO and 5RGx100, are meaningful as they generate positive values of q² > 0.6, 2-fold higher than the accepted value for a meaningful model (Cramer et al., 1993). To confirm that this model does not reflect a chance correlation, the response variables were permuted several times and cross-validation was repeated. In each case, after permuting the response variables, null or even negative q² values were obtained indicating that the models generated were not meaningful.

A prospective study using the models generated by both 3D-QSAR methods enabled the prediction of the K_{m (apparent)} of substrates in the test set. Catalyst succeeded in predicting arteether, verapamil, amitriptyline, and bupropion K_{m (apparent)} values as the residuals from the observed value were all less than one log unit. However, Catalyst failed in predicting lidocaine which was predicted as having a lower K_{m (apparent)} by 3.69 log units. A potential reason for this failure is that Catalyst considers a molecule active if it possesses or orients in the proper manner all of the essential features necessary for activity. Other factors could contribute to such an error such as a molecular volume which is not represented in the training set, leading to van der Waals interactions with the enzyme active site. The latter point would appear to be the case for lidocaine. This molecule was predicted to be very active because it was able to closely match all four pharmacophore features. However, the phenyl ring of lidocaine (Fig. 6) occupies a region in space that was not represented in the molecules present in the training set and may negatively interact with the active site of CYP2B6. This is a known limitation of Catalyst, and we can conclude that this molecule is therefore too dissimilar to the training set compounds in order to consider its prediction reliable using this technique.

The PLS MS-WHIM model was very successful in predicting lidocaine, arteether, amitriptyline, and bupropion with residuals lower than one log unit. PLS MS-WHIM poorly predicted one molecule, verapamil, which was predicted as having a lower K_{m (apparent)} by 3.7 log units. However, the Fisher's exact test highlighted that this compound is dissimilar to the training set compounds. The F-value is considerably higher with respect to the well predicted test set compounds (Table 3). Therefore, the PLS MS-WHIM-predicted K_{m (apparent)} value for verapamil was considered unreliable.

To our knowledge, Catalyst has not been previously compared with other 3D-QSAR techniques such as MS-WHIM. In this study, activity predictions obtained with Catalyst had larger residuals than the predictions generated from the PLS MS-WHIM model. This difference between the two 3D-QSAR models may be centered around reliance on the geometric disposition of pharmacophore features for Catalyst and our selection of the first of 10 hypotheses, rather than the multiple shape and electronic descriptors used in the PLS MS-WHIM technique. The first Catalyst hypothesis was considered the best as the second and subsequent hypotheses were closer to the null and therefore not considered. We are aware of the limited variance of the substrate K_{m (apparent)} values present in the test set (3.3-4.5 molar log units) and concentration around the mean (3.9 molar log units). This is close to the training set mean K_{m (apparent)} value for this model (4.1 molar log units), which covers a much larger range of K_{m (apparent)} values (1.7-5.9 molar log units). Further molecules need to be included in the test set to increase the range of activity and confidence in the predictive nature of the models. Even though the Catalyst pharmacophore model may be less quantitative than the PLS MS-WHIM model, it may be of qualitative value as it provides a visual representation of molecular features necessary for substrates to fit in the active site of CYP2B6. The reasoning behind our choice of 1 log unit residual as our cutoff for predictions was due to the variability of much of the published in vitro derived kinetic data, which can vary significantly (>1 log) between laboratories for the same biotransformation. Therefore, we feel that predicting within a 1 log unit residual and producing an approximate rank order of the activity may be as useful as predicting in vitro data with greater precision.

In order to prospectively generate a test set for our 3D-QSAR models the present study identified a number of substrates for B-lymphoblastoid expressed CYP2B6 which are also substrates for other human CYPs. The first example is the local anesthetic and antiarrhythmic lidocaine which has been widely recognized as a substrate for CYP3A4 in human liver microsomes. Recently, CYP2B6 has been implicated in lidocaine N-deethylation to form MEGX (Imaoka et al., 1996). The current report illustrated that CYP2B6 has a lower K_{m (apparent)} for lidocaine (537.6 µM) compared to that described for the CYP3A4 mediated high K_{m (apparent)} component (1388-1927 µM)(Bargetzi et al., 1989). The finding that amitriptyline is a substrate for lymphoblastoid-expressed CYP2B6 (Table 1) is not unexpected since the N-demethylation of another tricyclic antidepressant, imipramine, is catalyzed by CYP2B6 (Koyama et al., 1997). Interestingly, the K_{m (apparent)} for amitriptyline is approximately half the value reported for the K_{m (apparent)} of imipramine. The current study suggests that although both tricyclics differ in the saturation of only one aliphatic C-C bond and the addition of a nitrogen atom, this can have a marked effect on substrate affinity for CYP2B6. The physiological role of CYP2B6 with regard to amitriptyline N-demethylation is questionable since its K_m (144 µM) is substantially greater than that (20 µM) of an implicated unidentified CYP (Gharamani et al., 1997). The calcium channel antagonist, verapamil, has been reported to be metabolized to norverapamil by CYP3A with CYP1A2 also participating in vitro (Kroemer et al., 1992). Interestingly, the current study suggests that CYP2B6 is responsible for formation of O-desmethyl verapamil metabolite as opposed to the N-demethylated metabolite. Verapamil O-demethylation demonstrated non-Michaelis-Menten kinetics, which were then fitted to the Hill equation (Table 3), as described previously for 7-ethoxy-4-trifluoromethycoumarin (Ekins et al., 1997) and testosterone (Ekins et al., 1998). In agreement with a previous study, lymphoblastoid-expressed CYP2B6 was also the major CYP responsible for bupropion hydroxylation (Wurm et al., 1996), although no K_{m (apparent)} value had been generated previously.

Future work will involve further developing the Catalyst-produced CYP2B6 model. In addition, this same pharmacophore can also be used to search chemical databases to identify and obtain substrate probes for CYP2B6. These are all potential directions that would enable further understanding of CYP2B6 and are made possible by the 3D nature of the Catalyst-generated pharmacophore. In the meantime, the PLS MS-WHIM model is preferred to the Catalyst model, based on cross-validation procedures for the prediction of K_{m (apparent)} values for CYP2B6 substrates.