Materials and Methods

Molecular Modeling.

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

Modeling with Catalyst.

An extensive literature search was initiated to identify K_{m (apparent)} values for CYP3A4 substrates determined in incubations with human liver microsomes (Table 1). When multipleK_{m (apparent)} values for the same biotransformation were reported, the mean value was taken, and this is clearly indicated (Table 1). 3D structures of these substrates were built interactively using Catalyst Version 2.3 or 3.1. The number of conformers generated using the “best” feature of the program for each substrate was limited within the program to a maximum of 255 with an energy range of 20 kcal/mol. Ten hypotheses were generated using these conformers for the molecules in the training set and theK_{m (apparent)} values after selection of the following features: hydrogen bond donor, hydrogen bond acceptor, hydrophobic and negative ionizable. After an assessment of all 10 hypotheses generated, the lowest energy cost hypothesis was considered the best because this possessed features representative of all the hypotheses. The total energy cost of the generated pharmacophore hypothesis can be calculated from the deviation between the estimated activity and the observed activity, combined with the hypothesis complexity (feature number). A null hypothesis can also be calculated that presumes no relationship in the data and in which experimental activities are normally distributed about the mean value (Catalyst Tutorials Release 3.0). Therefore, the greater the difference between the energy cost of the generated and null hypotheses, the less likely it is that the generated hypothesis reflects a chance correlation.

The goodness of the structure-activity correlation was estimated by means of r. Molecular Simulations suggest validation of the retrieved hypothesis by permuting the response variable [i.e., the activities of the training set compounds were randomized a number of times so that each K_{m (apparent)} value was no longer assigned to the original molecule, and the Catalyst hypothesis generation procedure was repeated]. A typical cross-validation (q²) procedure does not appear to be recommended due to the nature of the hypothesis generation algorithm, which eliminates pharmacophore configurations shared by the most active and least active two molecules to leave a set of pharmacophore features able to discriminate active from nonactives (Molecular Simulations, Catalyst technology update seminar).

A subset of three substrates that are known to demonstrate autoactivation kinetics of human liver microsomal CYP3A4 (testosterone, nifedipine, and carbamazepine) were analyzed using the common features procedure within Catalyst, which produces an alignment independent of activity (Catalyst Tutorials Release 3.0). The resultant common features hypothesis was compared with the hypothesis previously derived from 38 molecules containing these three substrates. In addition, these models were compared with a hypothesis generated for 35 molecules, excluding the three substrates that demonstrate activation kinetics.

Prediction with Test Set Molecules.

As another form of pharmacophore validation, Molecular Simulations also suggest assessing whether the rank order of activities is accurately predicted for the training set and a test set of molecules. The test set of 12 molecules not included in the initial training set of 38 molecules were fit to the Catalyst hypothesis to predict the K_{m (apparent)} value using the fast fit and best fit feature of the program. “Fast fit” refers to the method of finding the optimum fit of the substrate to the hypothesis among all the conformers of the molecule without performing an energy minimization on the conformers of the molecule. This method is also used to fit the training set molecules to the pharmacophore. The “best fit” procedure starts with the fast fit procedure and allows individual conformers to flex over an energy threshold of 20 kcal/mol. This allows examination of more conformational space and minimizes the distance between the hypothesis features and the atoms to which they map (Catalyst Tutorials Release 3.0). The predictions using best and fast fits were compared by means of a residual that was calculated from the difference (in log units) between predicted and observed K_{m (apparent)} values. A predicted K_{m (apparent)} value within 1 log unit of the observedK_{m (apparent)} value was considered to be a valid prediction of fit because different laboratories frequently report K_{m (apparent)} values outside of this range (e.g., midazolam, as reviewed by Ekins et al., 1998b).

Results

Catalyst CYP3A4 Substrate Pharmacophore.

To define a pharmacophore, Catalyst uses the overall combination and 3D spatial distribution of structural features of the ligands and a measure of binding site properties of an enzyme, such as theK_{m (apparent)} (Nelsestuen and Martinez, 1997). The previously reported K_{m (apparent)}values for 38 substrates of human liver microsomal CYP3A4 (Table 1) enabled the construction of a pharmacophore for substrates of this enzyme (Fig. 1). Multiple conformers of all 38 substrates were used to generate this pharmacophore in which the K_{m (apparent)} values ranged over 4 orders of magnitude (Table 1). The lowest energy pharmacophore possessed four features necessary for substrates to interact with the active site of CYP3A4: two hydrogen bond acceptors, one hydrogen bond donor, and one hydrophobic region (Fig. 1). The interhydrogen bond acceptor distance was 7.7 Å, whereas the hydrophobe and hydrogen bond donor were 6.6 and 6.4 Å from one of the hydrogen bond acceptors, respectively (Fig. 1). These coordinates were approximately similar for all 10 generated hypotheses. The substrate saquinavir had the lowest K_{m (apparent)} value in the dataset for CYP3A4, 0.35 μM, and was found to fit to all four pharmacophore features (Fig. 2). The selected lowest energy pharmacophore also demonstrated a good correlation of observed versus estimated K_{m (apparent)} values using the fast fit function (r = .67).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

The interbond angles and distances between pharmacophore features for the CYP3A4 Catalyst model using 38 substrates. Shown are a hydrophobic area (left sphere, cyan), hydrogen bond donor (lower spheres, purple), and two hydrogen bond acceptor features (upper right and right spheres, green) with vectors in the direction of the putative hydrogen bond donors.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Saquinavir fitted to the CYP3A4 Catalyst pharmacophore produced from the data in Table 1. Shown are a hydrophobic area (left sphere, cyan), hydrogen bond donor (lower spheres, purple), and two hydrogen bond acceptor features (upper right and right spheres, green) with vectors in the direction of the putative hydrogen bond donors.

For the CYP3A4 substrate model generated here, the total energy cost of the hypothesis (arbitrary units) was 183.3, whereas the energy cost of the null hypothesis was 192.6 (Table 2) and the fixed cost of the ideal hypothesis was 147.1. To test whether the energy difference (9.3 units) between the hypothesis and null hypotheses was significant, the K_{m (apparent)} values were permuted to randomize the activity with the molecular structures. After permuting twice, Catalyst generated hypotheses with a different arrangement of pharmacophore features, a smaller difference between the generated hypothesis and the null hypothesis, as well as a decreased pharmacophore fit (r = .52 and .49; Table 2).

Test Set Predicted K_{m (apparent)} and Observed K_{m (apparent)}.

After construction of the Catalyst pharmacophore from the training set of 38 substrates, it was used to predict the K_{m (apparent)} values for a test set of 12 molecules not included in the training set. These predicted values were then compared with the corresponding observed values generated with human liver microsomes. Table 3 reports predicted and observedK_{m (apparent)} values for these biotransformations using the CYP3A4 substrate pharmacophore. Catalyst was able to predict all the K_{m (apparent)}values of 12 substrates with residuals lower than 1 log unit when the molecules were fitted to the model by the fast fit method (i.e., when all conformers of each substrate were examined). In contrast, fitting by the best fit method resulted in poor fits for omeprazole, simvastatin, onapristone, and morpholine-urea-Phe-Hphe-vinylsulfone in the CYP3A4 pharmacophore model, as the residuals were greater than 1 log unit. The best fit method overestimated the substrateK_{m (apparent)} values for 9 of the 12 the test set substrates of CYP3A4, versus 6 of 12 for the fast fit method.

Catalyst Common Features Model of CYP3A4 Substrates Known To Autoactivate Their Metabolism.

Because 3 substrates (carbamazepine, nifedipine, and testosterone) within the 38 used in the CYP3A4 pharmacophore training set are known to autoactivate their own metabolism (Kerr et al., 1994; Shaw et al., 1997), it was decided to model the common features of these molecules. A procedure within Catalyst produces hypotheses without reliance on the biological activity data of the molecules selected, unlike the normal Catalyst hypothesis generation. The goal in this case was to compare the common features Catalyst hypothesis of the substrates that autoactivate their metabolism with the original CYP3A4 model derived from all 38 molecules. In addition, Catalyst hypothesis generation was carried out for 35 of the original 38 molecules, without the 3 molecules used for common features analysis. The overall intention of this procedure was to recognize features present within substrates that demonstrate autoactivation and that might otherwise be hidden in such a large dataset. The common features pharmacophore (Fig.3) derived from the three substrates that demonstrate autoactivation illustrates three hydrophobic areas and one hydrogen bond acceptor. The hydrophobic areas are located at distances of 4.4 to 7.6 Å from the hydrogen bond acceptor feature (Fig. 3). When the three substrates are fitted to this pharmacophore, they are shown to closely overlap (Fig. 4). In addition, the sites of metabolism for testosterone and carbamazepine are co-located (Fig. 4). The removal of these three substrates from the original pharmacophore did not significantly affect either the number and type of pharmacophore features, the r value, or the hypothesis-null cost difference (Table 2).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

The interbond angles and distances between features for the Catalyst common features model built using three CYP3A4 substrates known to autoactivate their own metabolism (testosterone, nifedipine, and carbamazepine). Shown are three hydrophobic areas (three left spheres, cyan) and one hydrogen bond acceptor (right spheres, green) with vector in the direction of the putative hydrogen bond donor.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4

Testosterone (green), nifedipine (red), and carbamazepine (blue) fitted to the Catalyst common features model built using these three CYP3A4 substrates known to autoactivate their own metabolism. The sites of metabolism are indicated by arrows of the same color. The model shows the three hydrophobic areas (left three spheres, cyan) and one hydrogen bond acceptor (right spheres, green) with vector in the direction of the putative hydrogen bond donor.

Discussion

There have been many published examples of 3D-QSAR models of enzyme substrates and inhibitors (for a review, see Kubinyi, 1997), but as yet there has been no attempt to model requirements of substrates for the active site of CYP3A4 using a 3D-QSAR approach. A recent review series summarized the properties of CYP3A4, undoubtedly the most important mammalian CYP (Smith et al., 1997b). Smith and colleagues (1997a) detailed the CYP3A4 active site characteristics, as well as those of the other major mammalian CYPs, based on homology models built using soluble bacterial CYP structures as a template. They also suggested the binding of CYP3A4 substrates in the active site is due to lipophilic forces, based on evidence from octanol and cyclohexane partition coefficients (Smith et al., 1997b). In addition, the flexibility of the conformation of the CYP3A4 active site indicated by many researchers may contribute to the overall diverse substrate selectivity of CYP3A4 (Smith et al., 1997b).

Most of the information on the active site of CYP3A4 has been gleaned from homology models based on CYPcam (Ferenczy and Morris, 1989), CYPBM3 (Lewis et al., 1996), or all four known crystallized bacterial CYPs (Szklarz and Halpert, 1997). These studies when combined suggest there are a number of possible sites for interaction with substrates in the active site of CYP3A4. In parallel, other approaches have been successfully used to determine the catalytic nature of this CYP. The potential for flexibility in the conformation of CYP3A4 (multiple conformers) has been examined using carbon monoxide binding kinetics of CYP3A4 (Koley et al., 1995) as well as other CYPs. The work of Koley et al. (1997) also suggested that drug-drug interactions may occur with different conformers of CYP3A4. A further suggested feature of CYP3A4 is that its active site may simultaneously contain more than one compound (Shou et al., 1994), which may be interrelated with the suggested different conformers. Although we are unaware of the structural details of these multiple binding sites or multiple conformers, there has been some speculation as to how metabolism of both molecules may occur in this situation (Korzekwa et al., 1998). A limited number of site-directed mutagenesis studies with CYP3A4 have also been carried out to provide more information on the CYP3A4 active site. These studies have identified key amino acid residues localized to three substrate recognition sites that have a differential role in altering substrate specificity, protein structure, and flavonoid stimulation (Harlow and Halpert, 1997; He et al., 1997; Domanski et al., 1998). However, one drawback of these site directed-mutagenesis studies is the limited number of substrates and activators of CYP3A4 that were used to interact with the mutant CYPs (Harlow and Halpert, 1997; He et al., 1997; Domanski et al., 1998). Thus, additional studies are required to identify whether other amino acids have a role in altering substrate specificity, protein structure, or flavonoid stimulation when CYP3A4 is exposed to different substrates and activators.

In the current study, we used 38 substrates of CYP3A4 to produce the first 3D-QSAR model for substrates of this enzyme. The 3D pharmacophore hypothesized consists of four structural features: two hydrogen bond acceptors 7.7 Å apart, one hydrogen bond donor, and one hydrophobic region (Fig. 1). The presence of hydrogen bond acceptors in this CYP3A4 substrate pharmacophore agrees with the suggestion that hydrogen bonding of many substrates and inhibitors occurs with the Asn74 residue, identified by docking molecules in the homology modeled CYP3A4 active site (Lewis et al., 1996). This is opposite to the earlier theory put forward by Ferenczy and Morris (1989), who indicated that hydrogen bonds between the enzyme and substrate were not major interactions. Our model also confirms the findings of other groups because it includes a hydrophobic feature, implying an interaction with one or more of the suggested multiple hydrophobic domains in the active site of CYP3A4 as identified by alignment with bacterial CYPs (Ferenczy and Morris, 1989; Szklarz and Halpert, 1997).

We generated a Catalyst hypothesis for 38 CYP3A4 substrates that had a small cost difference between the total cost and the null hypothesis. After permuting, poorer-quality models were obtained. We therefore decided to investigate whether the original model was predictive; to do this, we used the Catalyst CYP3A4 substrate model to predict the observed K_{m (apparent)} of other substrates for CYP3A4 using both fast fit and best fit functions of Catalyst. We used a 1 log unit residual as a cutoff for predictions due to the variability of published in vitro data for the same biotransformation (Ekins et al., 1998b). The fast fit technique produced estimates with a less than 1 log unit residual for all 12 test set molecules. In contrast, there were four instances in which the best fit algorithm did not successfully predict the observed K_{m (apparent)}. The failure to predict the K_{m (apparent)} of the four substrates appears to be related to the fact that best fit examines the ability of a higher energy conformer to interact with CYP3A4, and this conformer may not be the physiologically relevant conformation because the energy is artificially high. Alternatively, in this best fit conformation there may be molecular features of the substrates that lead to unfavorable van der Waals interactions within the enzyme active site. Other groups have also previously recommended that the structures of test set compounds should not be too far removed from the training set to avoid this type of problem (Kubinyi, 1997). As an example of a poor prediction in this present study, morpholine-urea-Phe-Hphe-vinylsulfone was found to possess two phenyl rings that were outside of the four pharmacophore features when fit with best fit (Fig.5).

• Download figure
• Open in new tab
• Download powerpoint

Figure 5

Morpholine-urea-Phe-Hphe-vinylsulfone best fit to the Catalyst CYP3A4 38 substrate pharmacophore to demonstrate regions of the molecule that clearly do not fit the features, phenyl rings extending upwards (top) and backwards (bottom). Hydrophobic area (left sphere, cyan), hydrogen bond donor (lower spheres, purple), and two hydrogen bond acceptor features (upper right and right spheres, green) with vectors in the direction of the putative hydrogen bond donors are also shown.

After evaluating the variance of the test set of K_{m (apparent)} values (0.6–2.4 log units), we are satisfied that it was neither concentrated around the mean (1.7 log units) or similar to the training set mean K_{m (apparent)} value (2.4 log units). Importantly, the training set covers a slightly larger range of K_{m (apparent)} values (−0.4 to 3.7 log units) than the test set, illustrating that the test set predictions were interpolations from the training set. It may be useful in the future to assess how well the model can extrapolate beyond the training set to predictK_{m (apparent)} values. This would provide further details on the use of the CYP3A4 model. The iterative refinement of this pharmacophore by the addition of the test set to the training set could be used to account for the structurally different molecules that were originally poorly predicted. A refined model may be used in the future to generate better predictions of theK_{m (apparent)} values for CYP3A4 substrate. In the meantime, the present CYP3A4 model (Fig. 1) possesses pharmacophore features that fit within the structural requirements previously indicated for CYP3A4 substrates (Lewis et al., 1996). These requirements include a hydrogen bond acceptor atom 5.5 to 7.8 Å from the site of metabolism that is in turn 2 to 4 Å from the oxygen molecule in the heme (Lewis et al., 1996).

Because there have been many reports of substrates for CYP3A4 that illustrate autoactivation of their own metabolism (Ekins et al., 1998a), we decided to discern the common features of such substrates that were included in our training set. Fewer than 10% of the substrates in this training set had previously positively been shown to possess this property, including testosterone, nifedipine (Shaw et al., 1997), and carbamazepine (Kerr et al., 1994). When only the common features of these three substrates were modeled with Catalyst, we found that this new common features pharmacophore was markedly different from the Catalyst CYP3A4 pharmacophore of 38 substrates. The common features of testosterone, nifedipine, and carbamazepine include three hydrophobic regions and a single hydrogen bond acceptor (Fig. 3) in an arrangement unlike that for the complete 38-molecule training set of CYP3A4 substrates. We believe that CYP3A4 substrates that are known to autoactivate their own metabolism are more likely to have multiple hydrophobic features than are those that do not. Therefore, hydrophobic interactions with the CYP3A4 active site may be more important than hydrogen bonding for these same CYP3A4 substrates. Interestingly, the arrangement of features for CYP3A4 autoactivators is similar to the pharmacophore generated for CYP2B6 substrates, as the CYP2B6 Catalyst pharmacophore consists of three hydrophobic regions and a hydrogen bond acceptor for a training set that also includes autoactivators (Ekins et al., 1999). This finding suggests that the possession of multiple hydrophobic features is an important feature for autoactivators of other CYPs (Ekins et al., 1998a).

The combination of all 38 substrates produces a pharmacophore that suggests these substrates use hydrogen bonds when binding in the CYP3A4 active site. This does not change when testosterone, nifedipine, and carbamazepine are removed and the hypothesis is generated for 35 substrates instead of 38. The approach taken in the present study may also provide further evidence that there are two or more structurally separate regions responsible for substrate orientation and binding within the CYP3A4 active site. One or more of these regions appears to have a large hydrophobic domain that can accommodate substrates that demonstrate autoactivation. This domain could correspond to the “palisade” of aromatic residues suggested by Lewis et al. (1996) to occupy one side of the heme pocket, although this hypothesis requires further clarification. Alternatively, this hydrophobic domain could coincide with the regions identified within the three substrate recognition sites that have a role in flavonoid stimulation, as described above (Harlow and Halpert, 1997; He et al., 1997; Domanski et al., 1998 ). The Catalyst common features method as used in the present study may represent a unique approach to predicting and understanding substrates that demonstrate autoactivation. Our training set includes substrates structurally similar to the three autoactivators that have not been reported to show this behavior and may require further examination. The future use of this pharmacophore in a predictive context certainly deserves assessment.

In summary, in this study, we used a 3D-QSAR approach to model the structural requirements of substrates necessary for binding in the active site of CYP3A4 and then subsequent metabolism. It may be worthwhile to use the data from this study to attempt other 3D- or four-dimensional-QSAR approaches that would use further information contained in the molecular structures, such as descriptor-based models. Although we have previously used a descriptor-based approach to model other biochemical datasets (Bravi et al., 1997, Ekins et al., 1999), the majority of the present 38 molecules were not amenable to this technique due to a combination of their size and molecular flexibility, which is presently a limiting factor of such descriptor models. The selection of a subset of this present training set excluding these large flexible molecules will clearly be amenable to these types of descriptor-based analysis, although the amount of molecular diversity would be restricted. The 3D nature of the pharmacophore produced in this study suggests it may be useful in database searching to identify other CYP3A4 substrates. In addition, the molecular structures of the 38 molecules may be used to produce a receptor site model to clarify the features necessary for substrate binding in the CYP3A4 active site. All of these in silico techniques may eventually be used to predict aspects of drug metabolism that are usually performed in vitro or in vivo and that require the development of time-consuming assays unique to each new molecular entity. These in silico tools may also be used in a high throughput manner as a preliminary screening approach in pharmaceutical development.