Generation and Evaluation of a CYP2C9 Heteroactivation Pharmacophore

Heteroactivation of P450-mediated drug metabolism is a well known phenomena for CYP3A4, the major drug-metabolizing P450 isoform in humans (Benet et al., 1996), with extensive in vitro evidence supporting the idea of two or more binding sites for substrates, inhibitors, and heteroactivators within or near the active site (Shou et al., 1994; Ueng et al., 1997; Korzekwa et al., 1998; Domanski et al., 2000; Hosea et al., 2000; Kenworthy et al., 2001). Such non-Michaelis Menten kinetics presents a deviation from the assumptions traditionally made in quantitative prediction of in vivo drug clearance and drug interaction potential (Houston and Kenworthy, 2000). Apart from potentially introducing errors in in vitro-in vivo correlations, positive cooperativity of drug metabolism, leading to increases in affinity and/or catalytic efficiency, is a potential cause of drug interactions if translated to in vivo events (Egnell et al., 2003). Reports of atypical enzyme kinetics for other relevant drug-metabolizing P450 isoforms are rare and often only supported by single observations. Such scarce data on atypical kinetics have been reported for CYP3A5 (Korzekwa et al., 1998), CYP1A2 (Ekins et al., 1998), CYP2D6 (Kudo and Odomi, 1998), CYP2B6 (Korzekwa et al., 1998), CYP2C8 (Korzekwa et al., 1998), and to an increasing extent for CYP2C9 (Korzekwa et al., 1998; Hutzler et al., 2001, 2002, 2003). For CYP2C9, as for CYP3A4, experimental observations seem to indicate simultaneous binding of heteroactivator and substrate within or near the active site (Hutzler et al., 2001).

CYP2C9 is believed to be one of the more important P450s in drug metabolism and is involved in a number of clinical drug-drug interactions (Miners and Birkett, 1998). Several attempts at obtaining quantitative structure-activity relationship (QSAR) models for CYP2C9 ligand binding have been reported, reflecting the desire of early identification of CYP2C9 substrates and effectors in drug discovery. Although no universally predictive model is currently available, the presence of a hydrogen bond donor in the CYP2C9 ligand (Jones et al., 1993, 1996a), in many cases an anionic feature (Mancy et al., 1995), has been supported also by pharmacophore studies (Ekins et al., 2000) and is suggested to be situated approximately 7 to 8 Å from the site of metabolism (Mancy et al., 1995; Jones et al., 1996a). Furthermore, a hydrophobic zone between the site of metabolism and a potential cationic site on the protein has been suggested to be important for binding of the potent CYP2C9 inhibitor sulfaphenazole, a compound that is hypothesized to interact directly to the P450 heme iron via the aniline nitrogen (Mancy et al., 1996). The importance of this heme interaction has however been disputed based on tight binding analogs of sulfaphenzole not containing a basic nitrogen (Rao et al., 2000).

Few studies have addressed the potential common structural features conferring activation potency toward P450s. The only pharmacophore attempt to our knowledge was based on three autoactivators of CYP3A4 (compounds displaying sigmoidal velocity versus substrate curves, supposedly resulting from simultaneous, cooperative binding in the active site leading to conformational change) (Ekins et al., 1999). Furthermore, one attempt has been made to study structural requirements for CYP2C9 heteroactivation, by use of structural derivatives of dapsone, a heteroactivator of several CYP2C9 substrates (Hutzler et al., 2002). Electronic effects were suggested to be of importance for heteroactivation, because changing one of the electron donating para amino groups of dapsone to a nitro group, changed the effect from activation to inhibition (Hutzler et al., 2002).

The main purpose of this work was to test the hypothesis of common pharmacophore features within a diverse set of CYP2C9 heteroactivators and to compare heteroactivator pharmacophore features with those of inhibitors. We report the identification of previously unknown heteroactivators of CYP2C9 by use of a high throughput P450 effector screen.

Materials and Methods

Materials. Previously characterized (Masimirembwa et al., 1999) recombinant human cytochrome P450 (rCYP) 2C9 expressed in yeast was from AstraZeneca Biotech Laboratory (Södertälje, Sweden). 7-Methoxy-4-trifluoromethyl-coumarin (MFC) was from Sigma-Aldrich (St. Louis, MO), hydroxypropyl-β-cyclodextrin was from Aldrich Chemical Co. (Steinheim, Germany), dimethyl sulfoxide was from Lab-Scan (Dublin, Ireland), and black Greiner 384-well plates were from Greiner Labortechnik (Frickenhausen, Germany).

CYP2C9 High Throughput Effector Screening. The effect of a set of 1,504 compounds (oral drugs, known P450 inhibitors, and compounds chosen for maximum diversity using Diverse Solutions, http://www.tripos.com/sciTech/inSilicoDisc/comboChem/dvs.html) on CYP2C9-mediated metabolism of MFC to 7-hydroxy-4-trifluoromethyl-coumarin (HFC) (Fig. 1) was investigated using recombinantly expressed CYP2C9. Reaction conditions were as described previously (Bapiro et al., 2001) except for some modifications to allow for a high throughput approach. Briefly, compounds were dissolved at a concentration of 10 mM in a 1:1 mix of dimethyl sulfoxide and 50 mM cyclodextrin and were serially diluted in one-third increments by an automatic liquid handling system (CyBi-Well; Cybio Northern Europe Ltd., Jena, Germany) to yield seven different stock concentrations. One microliter of each stock concentration was transferred by the automatic liquid handling system, to the reaction plate and left at room temperature until evaporation of dimethyl sulfoxide was complete. A mix of rCYP2C9, substrate, and buffer was added to each well to yield final concentrations of 30 pmol/ml rCYP2C9, 50 μM MFC, and 0.025 μM KPO₄ (pH 7.4). Final organic solvent concentration from addition of substrate from a stock solution was 0.3% acetonitrile. After a 10-min preincubation at 37°C, reactions were started by the addition of NADPH (final concentration 1 mM). Metabolite formation was measured after 35-min incubation at 37°C by fluorescence measurement (excitation 405 nm, emission 535 nm) in a Spectra Max Gemini plate reader (Molecular Devices Corp., Sunnyvale, CA). Metabolite formation was linear with respect to time and protein concentration under the conditions used. To ensure that compounds tested were not fluorescent at the wavelengths used, the fluorescent emission of the compounds was measured in buffer. For heteroactivators, v_150% values (concentration resulting in 150% of control reaction velocity) were deduced from the plot of percent activity of control versus effector concentration by optical examination. Inhibitors identified in the effector screen were used in part of the evaluation of the pharmacophore. IC₅₀ values (concentration resulting in 50% inhibition) were calculated by fitting the data (using XLfit) to the following equation: where A is the minimum percentage of inhibition and B is the maximum percentage of inhibition. In cases where there were not enough data points to characterize the entire curve, B was set to 100 and A was set to 0.

Rationalizing the Use ofv_150% as Pharmacophore Modeling Parameter. The experimental endpoint used for building the heteroactivator pharmacophore was the concentration yielding 150% of control reaction velocity, v_150%. Due to the lack of detailed mechanistic knowledge about the molecular interactions leading to heteroactivation of this P450 isoform, an experimental parameter reflecting pure affinity (in analogy with IC₅₀ or K_i for competitive inhibitors) is not possible to derive. A theoretical relation between v_150% and the heteroactivator affinity constant K_a, can be made if making an analogy to the enzyme kinetic approach that has been used for CYP3A4 (Houston and Kenworthy, 2000; Kenworthy et al., 2001; Galetin et al., 2002). If assuming CYP2C9-mediated MFC metabolism and its heteroactivation can be described by parameters V_max, the maximum velocity of the reaction in absence of heteroactivation; S, the substrate concentration; A, the heteroactivator concentration; K_s, the affinity constant for binding of MFC; K_a, the affinity constant for binding of the heteroactivator to an hypothesized allosteric site; α, a value between 0 and 1 defining the change in affinity for the metabolic site induced by the heteroactivator (α < 1 for activation); and β, a value ≥0 defining the change in catalytical efficiency of HFC-epoxide formation induced by the heteroactivator (β >1 for activation), then the velocity in presence of a heteroactivator (v_activation) can be described as a function of V_max, S, A, K_a, K_s, α, and β, and the velocity in absence of a heteroactivator (v_control) can be described as a function of V_max, S, and K_s. At a fixed increase in velocity, [v_activation/v_control] takes a nominal value, for example for the modeling parameter used here, v_150%, [v_activation/v_control] = 1.5. Heteroactivator affinity K_a can then be expressed as a function of v_150%, V_max, S, K_s, K_a, α, and β. Because substrate concentration S was held constant between experiments, and because parameters V_max and K_s can be regarded as constants describing the interaction between MFC and CYP2C9 in absence of heteroactivator, it follows that if assuming that all heteroactivators are characterized by the same α and β, any change in K_a between heteroactivators would be reflected by a change in v_150%. The quantitative relationship between K_a and v_150% will differ depending on the mechanistic assumptions used to describe the interaction between the P450 and the substrate and heteroactivator, but if assuming a simplistic case in which all CYP2C9 heteroactivators have the same maximal potency of changing substrate affinity (i.e., all heteroactivators have the same α) and catalytic efficiency (i.e., all heteroactivators have the same β) v_150% would change proportionally with K_a and as such be a pure surrogate measure for ranking heteroactivator affinity.

The endpoint used to classify inhibitors in potency categories for validation of the heteroactivation pharmacophore was the concentration resulting in 50% inhibition, the IC₅₀ value, a pure surrogate measure for ranking inhibitor affinity if assuming competitive inhibition (Todhunter, 1979).

Molecular Modeling and Pharmacophore Generation. Three-dimensional (3D) structures of CYP2C9 MFC-heteroactivators were generated in CONCORD (Tripos Inc., St. Louis, MO) and imported into Catalyst (Accelrys, Princeton, NJ) where all subsequent modeling and pharmacophore generation was performed. For chiral compounds where racemates were used to generate experimental data, all possible isomers were included in the training set and assigned the same v_150% value. Three-dimensional minimization and conformer generation were performed within Catalyst, which uses a forcefield with a parameter set derived from the CHARMm force field. For each structure, a maximum of 255 conformers were generated using the “Best Quality” feature to maximize the coverage of conformational space. The generation of conformers in Catalyst is performed using the poling method (Smellie et al., 1995a,b,c), designed to cover as broad low energy conformational space as possible (Barnum et al., 1996), by introducing a so called poling function in the forcefield calculation to favor generation of conformers from previously unexplored conformational space.

Hypothesis generation was performed using all possible chemical features of the compounds in the training set as possible features of the pharmacophore. Features included were hydrogen bond acceptors, hydrogen bond donors, hydrophobes, negative ionizable features, and aromatic rings.

All molecular and pharmacophore modeling was performed on a Silicon Graphics O₂ station (Silicon Graphics Inc., Mountain View, CA).

Pharmacophore Validation. To validate that the quantitative heteroactivation pharmacophore chosen for further validation was not generated by chance, the pharmacophore generation was repeated using identical conditions except for permutation of v_150% values. Permutation was performed by randomizing values, making sure no compound retained its original v_150% value. Comparison of significance between pharmacophores was judged by the difference in cost between the lowest cost hypothesis and the related null hypothesis, as well as on correlation coefficients.

For external validation, five different modified training-sets were used, four different structures being left out in each, resulting always in leaving out one structure from within potency categories v_150% ≤ 1 μM, 1 to 20 μM, 20 to 60 μM, and >60 μM. Removal of heteroactivators from the training-set was performed by random within these predefined potency groups.

Due to the lack of potent heteroactivators for validating external predictivity, and to compare heteroactivator affinity features with those of inhibitors, a set of CYP2C9 inhibitors identified from the same experimental data set were tested for fit to the CYP2C9 heteroactivation pharmacophore. Inhibitors were chosen without prior knowledge of the structure, to cover defined IC₅₀ ranges (IC₅₀ < 1 μM, n = 52; 1–10 μM, n = 32; >10 μM, n = 124). A set of nonactives (no effect at 125 μM, n = 49) were included in the evaluation. The 3D structures of the compounds for inhibitors and nonactives were generated in CONCORD and imported in Catalyst for conformer generation as described above for heteroactivators. Compounds were fit to the CYP2C9 heteroactivator pharmacophore using the Fast Fit feature of the software.

Results

CYP2C9 Effector Screening. Of the 1504 compounds tested in the CYP2C9 high throughput effector screen, 37 (2.5%) were identified as heteroactivators of CYP2C9-mediated HFC formation (>20% activation at ≤125 μM). Heteroactivator structures and their respective observed v_150% values are shown in Table 1. Representative experimental data are shown in Fig. 2 for the four most potent heteroactivators (v_150% ≤ 1 μM). The drop in activation potency observed at high concentrations of some compounds, possibly due to solubility problems at high effector concentrations, did not affect the estimation of v_150%. Although the absence of a common level of maximum activation by all heteroactivators (Fig. 2) could theoretically be due to solubility problems, it cannot be excluded that the effect reflects a true difference in activation potency, in which case the modeling parameter, v_150%, would not reflect pure affinity but would contain information also about potency. Further insight into the mechanism of heteroactivation of CYP2C9 would be needed to allow derivation and in vitro measurement of an experimental parameter purely reflecting heteroactivator affinity.

Several of the most potent heteroactivators of MFC metabolism are oral drugs and endogenous compounds, including in decreasing potency amiodarone (antiarrythmic), niclosamide (antiparasitic), liothyronine (thyroidal hormone), meclofenemate (nonsteroidal anti-inflammatory drug), zafirlukast (antiasthmatic), estropipate (estrogen replacement), dichlorphenamide (carbonic anhydrase inhibitor), hydroflumethiazide (diuretic), mefenamic acid (nonsteroidal anti-inflammatory drug), and warfarin (anticoagulant), all with v_150% values below 10 μM. Of these compounds, only niclosamide has been previously reported as a CYP2C9 heteroactivator (Bapiro et al., 2001).

CYP2C9/MFC Heteroactivation Pharmacophore. Initially, two different training sets were used for pharmacophore generation, one excluding the structurally deviating compound 5 (Table 1), to investigate the impact of this compound on the final pharamacophore. Excluding compound 5 gave slightly better estimates of its prediction (data not shown), a higher r value for internal prediction of training-set compounds, as well as a slightly higher difference in cost to the corresponding null hypothesis (Table 2). The next highest scored pharmacophore when including compound 5 had identical (Table 2) and superimposable (not shown) features to the highest scored hypothesis generated when excluding this compound. The highest scored pharmacophore generated when excluding compound 5 was taken further for validation of statistical validity and predictive power. There was a significant difference in features, cost, and correlation coefficients of the heteroactivation pharmacophore compared with the lowest cost pharmacophore based on permuted v_150% values (Table 2), suggesting that the pharmacophore is a statistically valid representation of CYP2C9 heteroactivation of HFC formation. The geometric details of the pharmacophore is shown in Fig. 3, A and B, and superimposed with the five most potent heteroactivators in Fig. 4.

Of the training set compounds, 94% (34 of 36) were predicted within 1 log unit of the residual (r [log observed v_150%] versus [log predicted v_150%] = 0.71; r² 0.50). The categorical nature of the model, as suggested by Fig. 5A, is supported by the results of external validation attempts. The lowest cost external validation pharmacophore models (Table 3) generated by compound exclusion from the training set (Table 4) were all similar to the pharmacophore based on the complete training set (n = 36). The hydrogen bond acceptor/aromatic ring feature is conserved and in each case superimposable with the pharmacophore in Fig. 3 (not shown). Only four of five external validation models predicted the v_150% within 1 log residual for three of four compounds, with only one model (model 4) correctly ranking all compounds (Table 4). Due to the skewed distribution of v_150% values in the training set, using a default v_150% uncertainty value of 3 (defined by Catalyst as the measured value being within 3 times higher or 3 times lower of the true value), only the two most potent molecules in the training set are classified by Catalyst as actives to be used in the initial step of pharmacophore generation. As such, exclusion models 2 and 4 (Table 4) involve excluding one of the two compounds on which the original model was heavily dependent. The potent compounds excluded in models 3 and 5 were outliers also when included in the original model. In model 1, where both potent heteroactivators are kept in the training set, the two excluded potent heteroactivators are well predicted by both the original model and the external validation model. Model 1 does however generate a false positive (Table 4).

The ability of the CYP2C9 heteroactivation pharmacophore (Fig. 3) to classify potent inhibitors as tight CYP2C9 binders is shown in Fig. 5B. A high percentage of the most potent heteroactivators (Fig. 5A) and inhibitors (Fig. 5B) are predicted in the <1 μM range, whereas there is a steady decrease in this percentage as IC₅₀ increases. Few false hits (4%) are seen for nonactives (Fig. 5B).

Discussion

One limiting factor in structural investigations of CYP2C9 heteroactivators is the scarcity of experimental data on which to base computational methods. In an effort to overcome this limitation, we have used a high throughput approach to screen a large number of compounds (>1500) from which we have been able to identify 36 previously unknown heteroactivators of CYP2C9-mediated HFC formation, with v_150% values in the range 0.04 to 150 μM (Table 1).

Although several of the most potent heteroactivators identified are marketed drugs or endogenous compounds (Table 1), the possible clinical relevance of the findings is difficult to gauge, because our data, in combination with others (Hutzler et al., 2002), suggests that CYP2C9 heteroactivation is strongly substrate-dependent. This phenomena has been extensively studied with CYP3A4 (Kenworthy et al., 1999; Wang et al., 2000; Kenworthy et al., 2001), for which the substrate-dependent effects have been rationalized as a result of multiple and cooperative binding sites within or near the active site, as supported by enzyme kinetic modeling, spectral binding studies and site-directed mutagenesis (Shou et al., 1994; Ueng et al., 1997; Korzekwa et al., 1998; Domanski et al., 2000; Hosea et al., 2000; Kenworthy et al., 2001; Dabrowski et al., 2002; Galetin et al., 2002). Two clear examples of substrate dependence with CYP2C9 are with effectors amiodarone and zafirlukast. Both are potent heteroactivators of MFC metabolism (v_150% = 0.04 and 1.2 μM, respectively; Table 1) but are reported to be CYP2C9 inhibitors of other substrates (amiodarone/warfarin; zafirlukast/warfarin; and zafirlukast/tolbutamide) (O'Reilly et al., 1987; Heimark et al., 1992; Suttle et al., 1997; Vargo et al., 1997; Shader et al., 1999). Similarly, niclosamide, a potent activator of MFC metabolism (Bapiro et al., 2001) (Table 1), is a potent inhibitor of CYP2C9-mediated diclofenac-4-hydroxylase activity (Bapiro et al., 2001). Furthermore, dapsone, a heteroactivator of CYP2C9-mediated metabolism of flurbiprofene, naproxene, and piroxicam (Korzekwa et al., 1998; Hutzler et al., 2001) was observed here to be a weak inhibitor of MFC metabolism (IC₅₀ = 145 μM; data not shown). Another example is sulfamethoxazole, not studied in this work but reported by others to weakly inhibit tolbutamide metabolism (Back et al., 1988) but to have no effect on flurbiprofene or naproxene metabolism (Hutzler et al., 2002). Due to these evident substrate-dependent effects, direct extrapolation of the results presented here to other, potentially more clinically relevant substrates may not be valid. Furthermore, for clinical relevance to be assessed, a number of pharmacokinetic factors must be considered, such as therapeutic concentrations in relation to heteroactivation potency, hepatic extraction ratio, and administration route. P450 activity is expected to influence systemic drug clearance only when the extraction ratio of the affected drug is relatively low and could potentially influence presystemic clearance to lower the bioavailability of an orally administered drug of any extraction ratio (Pang and Rowland, 1977; Egnell et al., 2003). For example, even if the potent heteroactivating effect on MFC metabolism by niclosamide and zafirlukast (Table 1) was also relevant for other substrates, these drugs would not be expected to increase hepatic in vivo CYP2C9 activity because of very low systemic availability (Tracy and Webster, 1996) and low plasma concentrations (Shader et al., 1999), respectively.

Many of the heteroactivators are also known substrates for CYP2C9 (such as mefenemic acid, v_150% 5 μM; warfarin v_150% 5.5 μM; flurbiprofene v_150% 35 μM; and alprazolam v_150% 42 μM) (Leemann et al., 1992; Takahashi et al., 1998; Venkatakrishnan et al., 1998; Hutzler et al., 2002), further supporting the possibility of different binding modes and/or binding sites.

The main purpose of this work was to investigate whether CYP2C9 heteroactivator affinity can be related to structure, information that would be potentially useful for identification at an early stage of drug discovery. The generated pharmacophore (Fig. 3), based on 36 heteroactivators (v_150% 0.04–150 μM), was a significantly better representation of data than a pharmacophore based on permuted v_150% values (Table 2). The features defined by the pharmacophore, i.e., a hydrogen bond acceptor, an aromatic ring, and two hydrophobes, are in accordance with structural features previously suggested to be important for CYP2C9 active site binding. For CYP2C9 inhibitors, the presence of at least one hydrogen bond acceptor and one hydrophobic region, separated by 3 to 5.8 Å, has been suggested (Ekins et al., 2000). For these inhibition pharmacophores, aromatic rings were not included as possible features, and it seems possible that the hydrogen bond acceptor-aromatic ring feature of the heteroactivation pharmacophore, separated by 4.8 Å (Fig. 3A), could correspond to the hydrogen bond acceptor-hydrophobe feature in the reported inhibition pharmacophore (Ekins et al., 2000). The importance of hydrogen bonding and hydrophobic interactions for CYP2C9 binding is further supported by CYP2C9 homology modeling and the reported correspondence of active site amino residues to grid interaction energies in a 3D QSAR CYP2C9 inhibition model (Afzelius et al., 2001), as well as by an inverse pharmacophore based on analysis of selective regions in the active sites of different human CYP2C isoforms (Ridderström et al., 2001). The presence of an aromatic ring in the heteroactivation pharmacophore is in accordance with previous suggestions of the importance of aromatic stacking interactions for CYP2C9 binding, as based on a ligand based 3D QSAR combined with a CYP2C9 homology model-directed site-directed mutagenesis study showing the decrease in affinity of warfarin, diclofenac, and sulfaphenazole caused by a Phe114Leu mutation (Korzekwa and Jones, 1993; Jones et al., 1996b; Haining et al., 1999). An additional pharmacophore model for CYP2C9 inhibition (Rao et al., 2000) supports the pharmacophoric features in our heteroactivation pharmacophore in that it suggest the importance of aromatic stacking interactions in relative proximity to partially negatively and/or positively charged groups, as well as enhanced affinity with increased steric interactions further away from the aromatic ring, possibly corresponding to the additional hydrophobic features in the heteroactivation pharmacophore (Fig. 3).

The positive impact of halogenation of aromatic structures of benzbromarone derivatives on CYP2C9 affinity has been recently reported as based on inhibition of CYP2C9-mediated warfarin metabolism (Locuson et al., 2003). Electronic effects have been previously suggested to be of importance for CYP2C9 binding. In the case of a dapsone derivative, electron donation was suggested to be a determinant of whether the effect of the binder is activation or inhibition of flurbiprofene 4-hydroxylation (Hutzler et al., 2002), and similarly, compounds with similar steric bulk but different electrostatics varied greatly in their ability to inhibit S-warfarin metabolism (Rao et al., 2000). The importance of such effects for heteroactivation needs further experimental and computational investigation. The similarity in structure of the benzbromarone derivatives (Locuson et al., 2003) to amiodarone, the most potent heteroactivator in our training set, makes them interesting candidates for investigating structural and electronic effects of CYP2C9 binders on different marker substrates. The majority of these benzbromarone derivatives overlap with amiodarone to fit the hydrogen bond acceptor-aromatic ring feature of our heteroactivation pharmacophore as the lowest possible energy fit when using the Best Fit feature (data not shown).

Although quantitative internal predictivity of the presented CYP2C9 heteroactivation model (r of [log observed v_150%] versus [log predicted v_150%] = 0.71; r² 0.50) is in the range of that reported for CYP2C9 inhibition pharmacophores (Ekins et al., 2000), Fig. 5A suggests that the heteroactivation model is of categorical nature, with a higher degree of correct classification of high-affinity and low-affinity compounds than for intermediate potency compounds. One reason for the lack of continuous quantitative internal predictivity of the model might be the skewed distribution of the training set, the difference in v_150% of the most potent and the third most potent compound being more than 3.5 orders of magnitude. Because this causes Catalyst to use only the two most potent heteroactivators in the initial steps of pharmacophore generation, it is possible that the structural information is not optimally extracted. Increasing the number of compounds used by Catalyst in initial pharmacophore generation, by increasing the assumed uncertainty range of experimental values, yielded a hypothesis with similar features as that in Fig. 3, superimposable with the hydrogen bond acceptor-aromatic ring feature, but with higher cost than the null hypothesis (the theoretical cost for generating a hypothesis that contains no pharmacophore features and that estimates all activities to be the average activity) and lower internal predictivity (r [log observed v_150%] versus [log predicted v_150%] = 0.59; r² 0.35) than the model shown in Fig. 3 (data not shown).

Exclusion of compounds from the training set (Table 4) confirmed the expected lack of continuous quantitative predictivity of the model. Due to the scarcity of known potent CYP2C9 heteroactivators, the external categorical predictivity of such compounds using the model generated here cannot be properly tested. However, challenging the model with potent inhibitors as well as nonactives, all identified in the same high throughput screening assay, reveals a clear relationship between predicted and observed CYP2C9 affinity (Fig. 5B). It is also interesting to note that the most potent of the coumarin set of CYP2C9 inhibitors previously reported (Jones et al., 1996b) although not correctly classified (predicted affinity in the range of 14 μM for observed K_i ≤ 1 μM) fit well the hydrogen bond acceptor-aromatic ring feature of the heteroactivation pharmacophore presented here (Fig. 6).

The CYP2C9 heteroactivation model presented here, although not intended to be a high-resolution pharmacophore for CYP2C9 binding, still does reveal some important information. The clear trends in classification of CYP2C9 inhibitors (Fig. 5B) based on a model generated on heteroactivation data strongly indicates that potent heteroactivators and potent inhibitors do contain some common structural features conferring CYP2C9 active site binding, supporting previously reported experimental observations that positive cooperativity of CYP2C9 can be described by a model assuming two binding sites within the active site (Hutzler et al., 2001). Furthermore, the fact that important structural information on CYP2C9 affinity could be obtained based on high throughput screening heteroactivation data using a quick and simple pharmacophore approach based on a number of simplifying but controlled assumptions regarding affinity for active site binding is promising and suggests the future potential of quantitatively predictive heteroactivation QSAR models. It is tempting to speculate a future possible structure-based distinction between heteroactivators and inhibitors. Any generalization of such a distinction model would, however, probably be limited by the substrate-dependent effects.

In conclusion, by use of a high throughput approach, we have identified a number of previously unknown heteroactivators of CYP2C9-mediated MFC metabolism. The evident substrate-dependent effects suggest that it is advisable to use several substrates in CYP2C9 effector screens aimed at predicting clinical drug interaction potential. We propose that our data, in combination with others (Hutzler et al., 2001), suggest that potent heteroactivation of a screening marker substrate can be a potential sign of high CYP2C9 active site affinity, potentially causing inhibition with other substrates. The presented CYP2C9 heteroactivation pharmacophore is predictive for high active site affinity, suggesting that improvements of experimental and computational approaches may make possible the generation of continuously quantitative prediction models of CYP2C9 heteroactivator potency from chemical structure.