THREE- AND FOUR-DIMENSIONAL-QUANTITATIVE STRUCTURE ACTIVITYRELATIONSHIP (3D/4D-QSAR) ANALYSES OF CYP2C9 INHIBITORS

SEAN EKINS, GIANPAOLO BRAVI,1, SHELLY BINKLEY, JENNIFER S. GILLESPIE, BARBARA J. RING, JAMES H. WIKEL, AND

STEVEN A WRIGHTON

Departments of Drug Disposition (S.E., S.B., J.S.G., B.J.R., S.A.W.) and Computational Chemistry and Molecular Structure Research(G.B., J.H.W.), Lilly Research Laboratories, Eli Lilly and Co., Lilly Corporate Center, Indianapolis, Indiana

(Received January 31, 2000; accepted May 9, 2000)

This paper is available online at http://www.dmd.org

ABSTRACT:

The interaction of competitive type inhibitors with the active site ofcytochrome P450 (CYP) 2C9 has been predicted using three- andfour-dimensional quantitative structure activity relationship (3D-/4D-QSAR) models constructed using previously unreported andliterature-derived data. 3D-QSAR pharmacophore models of thecommon structural features of CYP2C9 inhibitors were built usingthe program Catalyst and compared with 3D- and 4D-QSAR partialleast-squares models, which use molecular surface-weighted ho-listic invariant molecular descriptors of the size and shape ofinhibitors. The Catalyst models generated from multiple conform-ers of competitive inhibitors of CYP2C9 activities contained atleast one hydrophobic and two hydrogen bond acceptor/donorregions. Catalyst model 1 was constructed with Ki(apparent) valuesfor inhibitors of tolbutamide and diclofenac 4*-hydroxylation (n 5

9). Catalyst model 2 was generated from literature Ki(apparent) values

for (S)-warfarin 7-hydroxylation (n 5 29), and Catalyst model 3 fromliterature IC50 values for tolbutamide 4-hydroxylation (n 5 13).These three models illustrated correlation values of observed andpredicted inhibition for CYP2C9 of r 5 0.91, 0.89, and 0.71, respec-tively. Catalyst pharmacophores generated with Ki(apparent) valueswere validated by predicting the Ki(apparent) value of a test set ofCYP2C9 inhibitors also derived from the literature (n 5 14). Twelveof fourteen of these Ki(apparent) values were predicted to be within 1log residual of the observed value using Catalyst model 1, whereasCatalyst model 2 predicted 10 of 14 Ki(apparent) values. The corre-sponding partial least-squares molecular surface-weighted holis-tic invariant molecular 3D- and 4D-QSAR models for all CYP2C9data sets yielded predictable models as assessed using cross-validation. These 3D- and 4D-QSAR models of CYP inhibition willaid in future prediction of drug-drug interactions.

The cytochrome P450s (CYPs)2 have been described as enzymeswith broad substrate specificity responsible for the metabolism ofmany drugs. Understanding the nature of the substrate specificity ofeach CYP requires knowledge of the interaction of drugs with eachCYP active site. However, due to the absence of a crystallized humanCYP, attempts to describe human CYP active sites have been re-stricted to homology modeling with the crystal structures of bacterialCYPs, pharmacophore, and three-dimensional-quantitative structureactivity relationship (3D-QSAR) studies (Smith et al., 1997a,b). Inparticular, 3D-QSAR studies have been very useful in rationalizingactive sites for many data sets of ligands or substrates for receptorsand enzymes (Kubinyi, 1997), suggesting their use for modelingCYPs as demonstrated recently (Ekins et al., 1999a,b,c,d).

CYP2C9 is one of the most important CYPs involved in drugmetabolism in humans (Miners and Birkett, 1998) and is involved inthe metabolism of many commonly used polar acidic drugs (Hall etal., 1994). CYP2C9 is also competitively inhibited by nonsteroidalanti-inflammatory drugs (Leeman et al., 1992; Newlands et al., 1992)as well as other classes of drugs (Schmider et al., 1997). Therefore,drug interactions with CYP2C9 are important in vivo and have beenreviewed (Miners and Birkett, 1998). The utility of predicting theability of a compound to inhibit CYP2C9 early in drug developmentwould therefore greatly increase the efficiency of this process. Therehas been a steady progression in understanding the active site ofCYP2C9 using a number of approaches. These studies have involvedusing overlapped substrates to define a binding template (Jones et al.,1993, 1996a). Others have used tienilic acid derivatives (Mancy et al.,1995), phenytoin analogs, and bis-triazole antifungals (Morsman etal., 1995) to build structure activity relationships to aid in understand-ing the substrate and inhibitor specificity of CYP2C9. In addition,site-directed mutagenesis has suggested the importance of the I-helixresidues Ser286 and Asn289 for conferring specificity for the sub-strates diclofenac and ibuprofen (Klose et al., 1998). NMR andmolecular modeling have also been combined to assist in defining thepositioning of substrates in the CYP2C9 active site (Poli-Scaife et al.,1997). This work suggested two structural determinants for substraterecognition, namely, an anionic site and a hydrophobic site (Poli-Scaife et al., 1997). Recently a CYP2C9 3D-QSAR comparative

1 Present address: Glaxo Wellcome Research and Development, MedicinesResearch Center, Gunnels Wood Road, Stevenage, Hertfordshire, SG1 2NY, UK.

2 Abbreviations used are: CYP, cytochrome P450; CoMFA, comparative mo-lecular field analysis; 3D-QSAR, 3 dimensional-quantitative structure activity re-lationship; 4D-QSAR, 4 dimensional-quantitative structure activity relationship;LOO, leave one out; MS-WHIM, molecular surface-weighted holistic invariantmolecular; PLS, partial least squares; 5RG3100, five random groups repeated upto 100 times.

Send reprint requests to: Sean Ekins, Ph.D., Department of Drug Disposition,Lilly Research Laboratories, Eli Lilly and Co., Lilly Corporate Center, Drop Code0730, Indianapolis, IN 46285. E-mail: ekins_sean@lilly.com

0090-9556/00/2808-0994–1002$03.00/0DRUG METABOLISM AND DISPOSITION Vol. 28, No. 8Copyright © 2000 by The American Society for Pharmacology and Experimental Therapeutics Printed in U.S.A.DMD 28:994–1002, 2000 /21/841508

994

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from

molecular field analysis (CoMFA) model was reported that describedcompounds that inhibited (S)-warfarin 7-hydroxylation and enabledleave one out (LOO) predictions ofKi (apparent)to be performed (Joneset al., 1996b). This model indicated the presence of at least one site ofelectrostatic interaction between the enzyme and substrate. This firstCYP2C9 CoMFA model enabled the refinement and construction of ahomology model of the CYP2C9 active site (Jones et al., 1996b). Veryrecently, site-directed mutagenesis studies have been combined withthe homology model to suggest the role of the amino acid residuesF114 and V113 in the hydrophobic binding of warfarin (Haining et al.,1999). However, the molecules used in the CoMFA study werestructurally similar and were also aligned manually. As with most3D-QSARs using CoMFA, the structural similarity and manual align-ment of the training set limit the usefulness of this CYP2C9 3D-QSAR. The LOO model in this case was also restricted to accuratepredictions for structurally similar molecules only, which is obviouslya disadvantage if the model were intended to predictKi values formolecules from other more diverse structural classes. Therefore, for apredictive 3D-QSAR, structural diversity is required but this comes atthe considerable cost of difficulty in alignment of common features ofmolecules.

Catalyst is a commercially available 3D-QSAR technique thatgenerates a representative set of conformers of molecules in a trainingset that accounts for the maximum occupation of conformationalspace of chemical functionalities (Quintana et al., 1995). Catalyst,unlike CoMFA, does not require manual alignment of molecules.Instead, the program generates a model from the chemical features ofthe appropriate conformers representing features involved in bindinginteractions with the enzyme after correlating measured and estimatedbiological activity. Previously, Catalyst was used to build substratepharmacophores for CYP2B6 (Ekins et al., 1999c) and CYP3A4(Ekins et al., 1999d), as well as inhibitor pharmacophores forCYP2D6 (Ekins et al., 1999a) and CYP3A4 (Ekins et al., 1999b). Inthis study, different QSAR modeling approaches, Catalyst and partialleast-squares (PLS) molecular surface-weighted holistic invariant mo-lecular (MS-WHIM) 3D- and 4D-QSARs (Klein and Hopfinger,1998) are compared for CYP2C9 using the in vitroKi(apparent)andIC50 values for inhibitors.

Experimental Procedures

Materials. Tolbutamide, diclofenac, chlorpropamide, triethylamine, andNADPH were obtained from Sigma Chemical Co. (St. Louis, MO). Chloro-form and acetonitrile were obtained from Burdick & Jackson Laboratories Inc.(Muskegan, MI). Meclofenamate was purchased from Cayman Chemical (AnnArbor, MI). 4-Hydroxy tolbutamide was purchased from Research Biochemi-cals International (Natick, MA). 49-Hydroxy diclofenac was obtained fromGentest Corp. (Woburn, MA) or Ultrafine Chemicals, (Manchester, UK).

Liver Specimens.Human liver specimens were obtained from the livertransplant unit at the Medical College of Wisconsin (Milwaukee) and thePathology Department of the Indiana School of Medicine (Indianapolis), underprotocols approved by the appropriate committee for the conduct of humanresearch. Microsomes were prepared from these specimens using differentialcentrifugation (van der Hoeven and Coon, 1974).

Microsomal Incubations. All incubations containing human liver micro-somes were carried out with 1 mM NADPH in 100 mM sodium phosphate pH7.4 and underwent a 3-min preincubation at 37°C. Inhibitors were used at fourconcentrations whereas the following substrates were used at five differentconcentrations.

Tolbutamide 4-Hydroxylation. Using a previously published method(Miners et al., 1988), 25–300mM tolbutamide was incubated with or withoutdiffering concentrations of inhibitor for 20 min and the reaction was terminatedby the addition of 20ml of 35% phosphoric acid. The internal standardchlorpropamide (0.01 mg/ml of stock) was added, and the incubation mixturewas extracted with 1 ml of chloroform. The vials were then mixed in a shaker

for 10 min and centrifuged. The aqueous layer was discarded while thechloroform layer was transferred to clean tubes and evaporated under nitrogenbefore reconstitution in mobile phase and injection of 50ml onto the HPLC.

The isocratic HPLC system used a mobile phase consisting of 50 mMpotassium phosphate (pH 3)/acetonitrile (73:27) and used a 5-mm SB-CNcolumn, 4.63 150 mm. The flow rate was 0.9 ml/min and the UV detector wasset at a wavelength of 230 nm.

Diclofenac 4*-Hydroxylation. Diclofenac (2.5–100mM) was incubatedwith or without differing concentrations of inhibitor for 15 min and terminatedby the addition of 200ml of acetonitrile and 10ml of internal standard,meclofenamate (0.05 mg/ml of stock). The tubes were centrifuged and 50mlof the supernatant was then injected on the HPLC.

The HPLC system used a mobile phase consisting of 50 mM sodiumphosphate buffer, pH 7.4, with 0.003% triethylamine (v/v) and used a betabasicC18 column 5mm (Keystone Scientific), 4.63 50 mm. The flow rate was 1ml/min and the UV detector was set at a wavelength of 282 nm.

Kinetic Analysis. TheKi values for inhibition of CYP2C9 were determinedusing nonlinear regression analysis as described in detail elsewhere (Ring etal., 1995).

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

Modeling with Catalyst. Briefly, three models were constructed usingCatalyst version 2.3 or 3.1 (Molecular Simulations, San Diego, CA). Catalystmodel 1 was constructed withKi(apparent)values reported in Table 1, model 2with Ki(apparent)values reported by Jones et al. (1996), and model 3 with IC50

values reported by Morsman et al. (1995). The 3D structures of CYP2C9inhibitors were built interactively as previously described for other CYPs(Ekins et al., 1999a,b,c,d). The number of conformers generated using the‘best’ functionality for each inhibitor was limited to a maximum number of255, with an energy range of 20 kcal/mol. Ten hypotheses were generatedusing these conformers for each of the molecules, and theKi(apparent)or IC50

values after selection of the following features for the inhibitors: hydrogenbond donor, hydrogen bond acceptor, hydrophobic, and negative ionizable.After assessing all 10 hypotheses generated, the lowest energy cost hypothesiswas considered the best, as this possessed features representative of all thehypotheses and had the lowest total cost. This procedure was followed for alltraining sets.

The goodness of the structure activity correlation between the estimated andobserved activity values was estimated by means of an r value. Statisticalsignificance of the retrieved hypothesis was verified by permuting (random-izing) the response variable, i.e., the activities of the training set compoundswere mixed a number of times (so that each value was no longer assigned tothe original molecule) and the Catalyst hypothesis generation procedure wasrepeated. The total energy cost of the generated pharmacophores can becalculated from the deviation between the estimated activity and the observedactivity, combined with the complexity of the hypothesis (i.e., the number ofpharmacophore features). A null hypothesis can also be calculated that pre-sumes that there is no relationship in the data and the experimental activitiesare normally distributed about their mean. The greater the difference between

TABLE 1

Competitive inhibitors of human microsomal tolbutamide 49-hydroxylation anddiclofenac 49-hydroxylation (CYP2C9)

The inhibitor constant was determined as described underExperimental Procedures.

Inhibitor Ki(apparent)a

mM

LY213829 3.56 0.4LY156735b 4.46 0.9LY300502b 8.06 0.4LY335979b 12.36 3.0Piroxicam 16.96 2.6Phenytoinc 17.16 1.5Clozapinec 31.16 2.2LY300164 33.86 1.6LY333531b 94.36 14.2

a Parameter estimate6 S.E. of the parameter estimate.b Diclofenac 49-hydroxylation.c Ring et al., 1996.

9953D- AND 4D-QSARS OF INHIBITORS FOR CYP2C9

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from

the energy cost of the generated hypothesis and the energy cost of the nullhypothesis, the less likely it is that the hypothesis reflects a chance correlation.

Validation of Catalyst CYP2C9 Pharmacophores Using a Test Set ofKi(apparent)Values. The test set ofKi(apparent)values for 14 molecules was notincluded in the initial training sets for either Catalyst model 1 or 2 that weregenerated withKi values. These test set molecules were fit by fast-fit andbest-fit methods to the Catalyst hypothesis for eachKi(apparent)model to predicta Ki(apparent) value as previously described for other CYPs (Ekins et al.,1999a,b,c,). Fast fit refers to the method of finding the optimum fit of thesubstrate to the hypothesis among all conformers of the molecule withoutperforming an energy minimization on the conformers of the molecule. Thebest-fit procedure starts with fast fit and allows individual conformers to flexover a given energy threshold of 20 kcal/mol. This allows examination of moreconformational space and minimizes the distance between the center of thehypothesis features and their mapping to atoms on the molecule. The result ofbest fit is a conformer whose energy is generally higher than the starting one.The conformer selected after this process was the one giving the prediction forbest fit (Catalyst tutorials release 3.0; MSI, San Diego, CA.). Both of thesepredictions were compared by means of a residual, which was calculated fromthe difference (in log units) between predicted and observedKi(apparent)values.A predictedKi(apparent)value within one log unit of the observedKi(apparent)

value was considered to be a valid prediction of fit.3D- and 4D-QSAR Modeling PLS MS-WHIM. Each molecule was coded

into a SMILES (simplified molecular input line entry system) string format(Weininger, 1988). Atomic 3D coordinates and Gasteiger Huckel charges weregenerated by CONCORD 3.2.1 (CONCORD user’s manual; Tripos Inc., St.Louis, MO). MS-WHIM descriptors were computed using the programEL3DMD (Bravi et al., 1997; Bravi and Wikel, 2000a,b) as previously de-scribed for other CYPs (Ekins et al., 1999a,b,c).

MS-WHIM descriptors are a set of statistical parameters that contain infor-mation about the structure of the molecules in terms of size, shape, symmetry,and distribution of molecular surface point coordinates after “weighted” cen-tering and principal component analysis (PCA). The following weights wereapplied: 1) unweighted value, 2) positive electrostatic potential, 3) negativeelectrostatic potential, 4) hydrogen bonding acceptor capacity, 5) donor ca-pacity, and 6) hydrophobicity, which yield a total of 72 descriptors (17 for eachweight; see Bravi et al., 1997 for details on MS-WHIM calculation).

To generate 4D-QSAR this whole procedure was repeated after first usingthe CATCONF program within Catalyst to produce multiple conformers of theinhibitors. For each molecule and for each descriptor the mean, the highest andlowest values, the range, and the S.D. over the conformations were computed.This resulted in a maximum of 504 descriptors (85 for each weight).

PLS (Wold et al., 1993) was applied to correlate MS-WHIM descriptorswith the observedKi(apparent)values. Because variance associated with differentdescriptors can be very different, MS-WHIM was autoscaled so as to assignunit variance to each descriptor. Activity data were transformed as log(1/Ki) orlog(1/IC50). The optimum number of components in each PLS model gener-ated was determined through two cross-validation procedures: 1) LOO and 2)five random groups repeated up to 100 times (5RG3100) (Baroni et al., 1993).Within the latter protocol, training set molecules are randomly assigned to fivegroups. Because results are strongly influenced by initial random assignment,the entire cross-validation procedure is repeated many times to achieve ameaningful statistical result. The predictive power of the PLS model wasevaluated by means of q2 and S.D. of error of prediction (Baroni et al., 1993)computed as follows:

q2 5 1 2 FOi~Yobserved2 Ypredicted!2/Oi~Yobserved2 Ymean!

2GS.D. of error of prediction5 FOi(Yobserved2 Ypredicted)

2/nG 1/ 2

wheren is the number of compounds. When using the 5RG3100 protocol thereported q2 and S.D. of error of prediction values represent mean values over100 cross-validation cycles. The reliability of the PLS model was furtherverified by permuting the response variable several times (as described forCatalyst) and repeating cross-validation. MS-WHIM weight selection (Bravi

and Wikel, 2000a,b) was carried out in an attempt to improve the PLS modelsand determine the most relevant contributions to the structure-activity corre-lation.

Results

Catalyst CYP2C9 Pharmacophore Models.Catalyst uses a col-lection of molecules with CYP2C9 inhibitory activity over a numberof orders of magnitude to enable construction of hypotheses to explainthe variability of inhibition with respect to the position of featurespresent in the inhibitors. TheKi values reported in Table 1 forinhibitors of tolbutamide 4-hydroxylation and diclofenac 49-hydroxy-lation vary over 27-fold. However, the structures in this data set arediverse (Fig. 1). The lowest energy pharmacophore constructed byCatalyst and referred to as model 1 produced four features suggestedto be necessary for inhibition of CYP2C9, namely, two hydrophobes,one hydrogen bond donor, and one hydrogen bond acceptor (Fig. 2,Table 2). The Catalyst model 1 demonstrated a good correlation ofobserved versus estimatedKi values (r5 0.91, Table 3). This initialCYP2C9 hypothesis possessed a total energy cost of 42.1 (arbitraryunits) whereas the energy cost of the null hypothesis was 33.9 (Table3).

Two previously published in vitro CYP2C9 inhibition data sets wereused in this study as a data source for development of 3D-QSAR models(Morsman et al., 1995; Jones et al., 1996b). TheKi values for CYP2C9using the 7-hydroxylation of (S)-warfarin as the CYP2C9 catalytic activ-ity cover 2 orders of magnitude but the molecules are structurally similar(Jones et al., 1996b) (Table 3). The lowest energy pharmacophore fromthis data set referred to as Catalyst model 2 produced four featuresnecessary for inhibition of CYP2C9, namely, one hydrophobe, two hy-drogen bond donors, and one hydrogen bond acceptor (Fig. 3, Table 4).The Catalyst model 2 demonstrated a good correlation of observed versusestimatedKi value (r5 0.89, Table 2). Catalyst model 2 possessed a totalenergy cost of 118.2 whereas the energy cost of the null hypothesis was130.5 (Table 3). The IC50 values for CYP2C9 inhibition using the4-hydroxylation of tolbutamide as the substrate probe, covered 1 order ofmagnitude and consisted of two structural classes, phenytoin and bis-triazole analogs (Morsman et al., 1995) (Table 3). The lowest energypharmacophore for this data set referred to as Catalyst model 3, producedthree features necessary for inhibition of CYP2C9, namely, one hydro-phobe and two hydrogen bond acceptors (Fig. 4, Table 5). Catalyst model3 demonstrated a reasonable correlation of observed versus estimatedIC50 (r 5 0.71, Table 3) and a higher total cost (59.9) than the nullhypothesis (47.9).

An accepted method to validate the Catalyst pharmacophores(Ekins et al., 1999a,b,c,d) involves permuting the activities with theircorresponding structures. This procedure should produce no signifi-cant model (decreased r value), altered pharmacophore features com-pared with the original hypothesis, and an energy cost value close toor greater than the null hypothesis. Following this procedure twice,Catalyst model 1 demonstrated total energy costs of 35.9 and 42.8 andthe pharmacophore fit had decreased, although the same features wereselected (Table 6). After permuting (as described above) Catalystmodel 2 there was no difference in the total energy cost (130.3) whencompared with the null hypothesis (130.5), the pharmacophore fea-tures changed, and the fit decreased. In contrast, following permutingCatalyst model 3, it demonstrated a higher hypothesis total cost (55)than the null hypothesis (47.9), and the number of pharmacophorefeatures decreased along with the correlation (Table 6).

Catalyst CYP2C9 Pharmacophore Validation Using a Test Setof Ki(apparent) Values. After constructing the Catalyst 3D-QSARmodels with a training set ofKi(apparent)values, models 1 and 2were used to predictKi(apparent)values of a test set of molecules

996 EKINS ET AL.

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from

(Table 7). PredictedKi(apparent) data were generated from bothmodels for this test set ofR-7-fluorowarfarin,S-7-fluorowarfarin,Rac-6-fluorowarfarin, Rac-6,7,8-trifluorowarfarin,S-warfarin al-cohol, ipriflavone, 7-hydroxyisoflavone, ticlopidine, lansoprazole,fluvoxamine, paroxetine, sertraline, omeprazole, desmethylcitalo-pram, and desmethylsertraline. These predictedKi(apparent) datawere then compared withKi(apparent)values reported in the litera-ture, which had been generated in human liver microsomes.Twelve of fourteen molecules were predicted within the 1 log unitresidual cutoff, using the fast fit function by Catalyst model 1,whereas 10 of 14 were predicted within the 1 log unit residualcutoff using the best fit function. For catalyst model 2, 10 of 14molecules were predicted within the 1 log unit residual cutoffusing the fast fit function, whereas 7 of 14 molecules were pre-dicted within the 1 log unit residual cutoff using the best fitmethod.

PLS MS-WHIM 3D- and 4D-QSAR Models of CYP2C9 Data.ThePLS MS-WHIM method initially used the alignment of single molec-ular conformations to produce a 3D-QSAR. However, by using mul-tiple conformers and, therefore, multiple alignments of molecules, thistechnique can be used as a 4D-QSAR. The PLS MS-WHIM model forsingle conformers, 3D-QSAR, of inhibitors of tolbutamide 4-hydroxy-lation and diclofenac 49-hydroxylation (Table 1) generated a LOO q2

value of 0.40. This value improved on inclusion of multiple conform-ers, 4D-QSAR, of inhibitors to give a LOO q2 value of 0.54 (Table 8).A significant 3D-QSAR model could not be generated for theCYP2C9Ki model derived from inhibition of (S)-warfarin 7-hydroxy-lation (Jones et al., 1996b) as the q2 value was less than 0.3. However,in this case 4D-QSAR was able to generate a q2 value of 0.64 withfive components consisting of the molecular surface properties ofnegative electrostatic potential, hydrogen bond acceptor, hydrogenbond donor, and hydrophobicity (Table 8). Using the more crediblecross-validation protocol of 5RG3100 (described earlier) also re-sulted in a significant 4D-QSAR model for the Jones et al. data set asassessed by the 5RG q2 value of 0.55. In contrast, a 3D-QSAR wasgenerated for CYP2C9 using the IC50 data derived from the inhibitionof 4-hydroxylation of tolbutamide (Morsman et al., 1995) with a LOO

q2 value of 0.5 with the molecular surface weights of positive elec-trostatic potential and hydrogen bond acceptor. The q2 value was notimproved by using 4D-QSAR of the inhibitors in this case, as nosignificant model could be generated. The S.D. estimate predictionsare directly related to the uncertainty of predictions and provide anestimate of confidence limits in the data. These values were compa-rable in the cases where both 3D- and 4D-QSARs could be generated.

Random permuting of the response variable a number of times (sothat the inhibition values no longer corresponded to the same mole-cules) was adopted to test the validity of the obtained models. Thisprocedure should not result in a predictive model. In each case, a q2

of zero or lower was obtained, indicating that permuting the responsevariable does not result in a predictive PLS MS-WHIM model. Thisapproach confirms that the cross-validated PLS MS-WHIM modelsobtained for CYP2C9 with a q2 . 0.3 are statistically valid.

Discussion

The investigation of substrate and inhibitor requirements forCYP2C9 has started to receive increased attention. Initial work sug-gested a schematic model for CYP2C9 binding that contained hydro-phobic and hydrogen bonding regions (Guengerich et al., 1986; Dis-telrath and Guengerich, 1987). The structural requirements ofCYP2C9 were initially speculated on by others (Smith and Jones,1992) and were followed by a preliminary active site model ofCYP2C9 constructed with seven substrates (Jones et al., 1993) [laterexpanded to eight substrates and one inhibitor (Jones et al., 1996a)].This preliminary model indicated the distance and angle between thesite of metabolism and the site of the hydrogen bond donor was 6.7 Åand 133°, respectively. Another group, modeling 12 CYP2C9 sub-strates, suggested the site of hydroxylation is 7.8 Å away from ananionic site, which is in turn 4 Å from a cationic site on the protein,separated by an angle of 82.3° (Mancy et al., 1995). Both of theseactive site models have six compounds in common and used low-energy conformers for model building derived from different soft-ware, which might account for some of the dissimilarities in the finalmodels (Mancy et al., 1995; Jones et al., 1996a). A more recentCYP2C9 active site model using 10 molecules indicated a distance of

FIG. 1. Structures of competitive inhibitors used for generation of CYP2C9 Ki data and resultant CYP2C9 model 1 3D-QSAR.

9973D- AND 4D-QSARS OF INHIBITORS FOR CYP2C9

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from

6.5 Å from the site of metabolism to the hydrogen bond donor (Lewiset al., 1998). A study using bis-triazole and phenytoin analogs asinhibitors of CYP2C9 has also shown the importance of hydrogenbonding for potency (Morsman et al., 1995). There have been no

reports of models being constructed with this data, neither have therebeen any details published on distances and angles between chemicalfeatures published for CYP2C9 inhibitors. This study demonstratesthe use of CYP2C9 3D- and 4D-QSAR inhibitor models. The

FIG. 2. Catalyst model 1 produced from the data in Table 1 illustrating hydrophobic areas (cyan), hydrogen bond donor (HBD, purple), and a hydrogen bondacceptor feature (HBA, green) with a vector in the direction of the putative hydrogen bond donor.

TABLE 2

Three-dimensional coordinates of catalyst pharmacophore features for model 1

HBA 5 hydrogen bond acceptor; HBD5 hydrogen bond donor; H5 hydrophobe.

Coordinates

Pharmacophore Features

HBA HBAVector HBD HBD

Vector H H

X 23.78 25.26 0.83 21.51 23.7 1.7Y 22.33 21.38 20.2 21.71 20.36 22.52Z 1.37 3.82 3.89 5.00 21.62 1.18

TABLE 3

Summary of sources of data used for building human hepatic CYP inhibitory pharmacophore, their features and fit to data using Catalyst

H 5 hydrophobe; HBA5 hydrogen bond acceptor; HBD5 hydrogen bond donor.

CYP2C9Catalyst Model

Number ofMolecules in Model Inhibition Range Type of

Inhibition Pharmacophore Features PharmacophoreFit

Hypothesis–Null Energy Cost

r

1a 9 3.5–94.3mM Ki 2 H, 1 HBD, 1 HBA 0.91 42.1–33.92b 29 0.1–50mM Ki 1 H, 2 HBD, 1 HBA 0.89 118.2–130.53c 13 12.9–250mM IC50 1 H, 2 HBA 0.71 59.9–47.9

a Data used for model construction from Table 1.b Data used for model construction from Jones et al., 1996b.c Data used for model construction from Morsman et al., 1995.

998 EKINS ET AL.

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from

CYP2C9 3D-QSAR models were used to satisfactorily predict theKi(apparent)values of molecules not present in the training set of themodels.

The data set in Table 1 derived from nine inhibitors of tolbutamideand diclofenac hydroxylations produced Catalyst model 1 with an rvalue of 0.91 with four pharmacophore features (Table 2). The smalldifference of 11.8 arbitrary units between the hypothesis and null costchanged in one of two cases of permuting. The correlation decreasedbut the pharmacophore features maintained were similar to the initialhypothesis. This suggests that Catalyst model 1 was suboptimal (Ta-ble 6). The Catalyst model 2 constructed from 29 structurally similarCYP2C9 inhibitors of (S)-warfarin 7-hydroxylation (Jones et al.,1996b) produced a model with an r value of 0.89 and four pharma-cophore features (Table 3). Also, in agreement with a previous report,there was at least one hydrogen bond donor and acceptor pharma-cophore feature in the Catalyst model of this data set (Table 3) (Joneset al., 1996b). The hypothesis total cost was less than the null costvalue indicative of a significant model. On permutation, Catalystmodel 2 illustrated almost identical total and null energy costs, de-creased correlations, and altered pharmacophore features (Table 6),suggesting that the initial Catalyst model was valid. Further literature

data sets of CYP2C9 inhibitors have also been published documentingIC50 values for inhibition of tienilic acid hydroxylation (Mancy et al.,1995) and tolbutamide hydroxylation (Morsman et al., 1995). TheMorsman et al. data set contained two classes of structural analogs,and it was decided to model it using the same criteria as applied to thetwo Ki(apparent)data sets. Catalyst model 3 constructed from this datawith phenytoin analogs and bis-triazole antifungals (Morsman et al.,1995) produced a good r value of 0.71 for three pharmacophorefeatures even though the hypothesis total cost was higher than the nullcost (59.9 and 47.9, respectively; Table 3). After permuting, the rvalue decreased to less than 0.3, the pharmacophore features changed,but the total cost value did not change significantly (Table 6), sug-gesting that the original model was valid. In summary, permutingresulted in a decrease in the r value for all of the models, suggestingthat this may be a useful test for model validity. In addition, all threeCatalyst CYP2C9 inhibitor pharmacophores possessed at least onehydrophobic region and a hydrogen bond acceptor, suggesting thatthese are common features of all CYP2C9 inhibitors.

Previous studies have shown that even with small hypothesis-nullhypothesis energy cost differences, valuable models of CYPs can beconstructed (Ekins et al., a,b,c,d). This is demonstrated by the predic-

FIG. 3. Catalyst model 2 produced from the published data set (Jones et al., 1996b), illustrating hydrophobic areas (cyan), hydrogen bond donor (HBD, purple), anda hydrogen bond acceptor feature (HBA, green) with a vector in the direction of the putative hydrogen bond donor.

TABLE 4

Three-dimensional coordinates of catalyst pharmacophore features for model 2

HBA 5 hydrogen bond acceptor; HBD5 hydrogen bond donor; H5 hydrophobe.

CoordinatesPharmacophore Features

HBA HBA Vector HBD1 HBD1 Vector HBD2 HBD2 Vector H

X 21.18 22.41 1.49 20.23 0.42 2.63 23.42Y 0.48 1.40 1.03 20.79 21.98 23.65 1.88Z 21.66 24.25 2.36 4.01 20.01 21.15 20.26

9993D- AND 4D-QSARS OF INHIBITORS FOR CYP2C9

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from

tion of Ki(apparent)values for a test set of compounds not included inthe training set for model building. In this study using the fast fitmethod to Catalyst model 1, predictions ofKi (apparent)were within 1log unit residual for 12 of 14 inhibitors. Using the best fit method, thenumber of molecules predicted within this cutoff decreased to 10 of14. Catalyst model 2 (data of Jones et al., 1996b) produced 10 of 14and 7 of 14 predictions within the 1 log unit cutoff when the fast fitand best fit methods, respectively, were used. Even though the train-ing set used to build Catalyst model 1 was relatively small (n 5 9), itappeared to have produced more realistic predictions than the Catalystmodel 2 built with the much larger data set (n5 29) of Jones et al.This is likely due to the structural diversity of the training data set formodel 1 and the varied structural nature of the test set. Althoughindicative of important features necessary to fit in the active site,Catalyst model 2 would be expected to be better at predicting theKi

value of molecules similar in structure to that of its training set.However, this may not be strictly true as fluorowarfarin analogs were,on the whole, poorly predicted (Table 7). To assess the suitability ofthe test set, we compared theKi(apparent)range (expressed as the log ofthe micromolarKi(apparent)value) for the test and training sets. The testsetKi(apparent)range (,0–1.82) was not concentrated around the meanKi(apparent)value (1.24). Similarly, theKi(apparent)range for Catalystmodel 1 (0.54–1.97) was not concentrated around the meanKi(apparent)

value (1.18). In this case, the meanKi(apparent)values of both test andtraining sets were similar. The range of the Catalyst model 2Ki(appar-

ent) values (21–1.84) was larger than the test set and Catalyst model1. The meanKi(apparent)value in this case (0.82) was smaller than themeanKi(apparent)values for Catalyst model 1 and the test set. Theseobservations suggest that the test set used for both models wassuitable for interpolating predictions from both training sets. Theassessment ofKi prediction appears to be a useful approach fordetermining the quality of these CYP2C9 Catalyst models and mayultimately be more valuable as a determinant of Catalyst 3D-QSARmodel quality than the cost difference between the hypothesis and nullhypothesis.

Using theKi(apparent)value data in Table 1, PLS MS-WHIM modelswere also constructed (Table 8). Due to the nature of this 3D- and4D-QSAR approach, there is no graphical output and the models are,therefore, solely assessed by means of the LOO q2 value. The 3D-QSAR derived using PLS MS-WHIM produced a LOO q2 value of 0.4for eight of nine inhibitors with the weights negative potential andhydrogen bond donor. One molecule (LY333531) was excluded, aseither its molecular size or flexibility exceeded the limit imposed byour PLS MS-WHIM method. In contrast, the 4D-QSAR generatedwith this same software illustrated an improvement in the q2 value to0.54 (Table 8). The q2 values represent predictive models using thecriteria of Cramer et al. (1993), who have suggested that a q2 valuegreater than 0.3 indicates that the model is meaningful (predictive)and a q2 of 1.0 is optimal. The CoMFA study of Jones et al. (1996b)used PLS to analyze the steric and electrostatic features of 27 of 29ligands along with the inverse logarithm of the binding affinity, whichresulted in a q2 value of 0.7. In this study, PLS MS-WHIM was usedto modelKi data for all 29 ligands reported in the Jones et al. CoMFAstudy. A significant 3D-QSAR was not produced for this data set (q2

, 0.3), but a significant 4D-QSAR was generated for the followingweights: negative potential, hydrogen bond acceptor, hydrogen bond

FIG. 4. Catalyst model 3 produced from the published data set (Morsman et al., 1995).

The pharmacophore demonstrates hydrophobic areas (cyan) and a hydrogen bond acceptor feature (HBA, green) with a vector in the direction of the putative hydrogenbond donor.

TABLE 5

Three-dimensional coordinates of catalyst pharmacophore features for model 3

HBA 5 hydrogen bond acceptor; HBD5 hydrogen bond donor; H5 hydrophobe.

CoordinatesPharmacophore Features

HBA HBA Vector HBA HBA Vector H

X 1.80 3.84 21.02 23.74 21.02Y 20.24 1.96 24.01 25.24 1.12Z 21.14 21.34 20.34 20.66 2.40

1000 EKINS ET AL.

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from

donor, and hydrophobicity. This yielded a 5RG q2 value of 0.64(Table 8), similar to that published for 3D-QSAR of Jones et al.(1999b). The same PLS MS-WHIM approach was used for the phe-nytoin analogs and bis-triazole antifungal CYP2C9 inhibitors reportedby Morsman et al. This model generated a LOO q2 value of 0.5 for3D-QSAR and a 5RG q2 value of 0.45 for 4D-QSAR using theweights positive potential and hydrogen bond acceptor (Table 8). Thedata from 4D-QSAR would suggest that using multiple conformationsfor these three data sets (Fig. 1) produce valid models as assessed byq2 validation with values.0.45 (Table 8). In all three cases, there is

also some degree of agreement between the PLS MS-WHIM featureweights and the features identified by Catalyst. In the future, it will beinteresting to evaluate the predictive nature of these PLS MS-WHIMmodels using the test set detailed in this study. A comparison of both3D- and 4D-QSAR versions of PLS MS-WHIM would also be idealto indicate how maximizing conformational space affects the predic-tions. Also, it may be useful to obtain a test set of IC50 values toenable quantitation of the predictive ability of the IC50 model de-scribed in this study.

The 3D arrangement of Catalyst pharmacophore features provedinteresting to evaluate (Tables 2, 4, and 5). Qualitatively, the generalshape of all three Catalyst CYP2C9 pharmacophores were similarwith the distances between a hydrogen bond acceptor and a secondhydrogen bond acceptor/donor being 3.4 to 5.7 Å (Figs. 2–4). Sec-ondly, a hydrophobic feature was positioned 3 to 5.8 Å from ahydrogen bond acceptor. All of these pharmacophores fit within thesubstrate template distances between residue contacts and sites ofmetabolism suggested for CYP2C9 modeled after alignment with thebacterial CYPBM-3 (Lewis et al., 1998). Among all three CYP2C9substrate models previously reported (Jones et al., 1993, 1996a; Lewiset al., 1998) there is little difference in the diversity of molecules used.This is unlike the present study in which three models were con-structed from unique data sets composed of a range of structurallydiverse CYP2C9 inhibitors. These multiple data sets were importantas they helped confirm multiple interaction determinants as necessaryfor inhibitors (hydrophobic and hydrogen bond acceptor/donor fea-

TABLE 6

Summary of permuted Catalyst hypotheses

H 5 hydrophobe; HBA5 hydrogen bond acceptor; HBD5 hydrogen bond donor.

CYP2C9CatalystModel

Permuted Hypothesis 1 Permuted Hypothesis 2

Pharmacophore Features Hypothesis–NullEnergy Cost

PharmacophoreFit Pharmacophore Features Hypothesis–Null

Energy CostPharmacophore

Fit

r r

1a 1 HBA, 1 HBD, 2 H 35.9–33.9 0.81 1 HBA, 1 HBD, 2 H 42.8–33.9 0.692b 1 HBA, 1 HBD, 2 H 130.3–130.5 0.58 1 HBA, 3H 130.4–130.5 0.663c 2 HBA 55.4–47.9 0.29 2 HBA 55–47.9 0.20

a Data used for model construction from Table 1.b Data used for model construction from Jones et al., 1996b.c Data used for model construction from Morsman et al., 1995.

TABLE 7

Observed and predictedK i (apparent) for inhibitors of CYP2C9 fit to different Catalyst CYP2C9 pharmacophores

Values in parentheses represent the residual (log units) of predicted2 observedKi (apparent)values.

Inhibitor ReferenceObserved

MicrosomalKi(apparent)

Model 1a

Prediction FastFit Ki(apparent)

Model 1a

Prediction BestFit Ki(apparent)

Model 2b

Prediction FastFit Ki(apparent)

Model 2b

Prediction BestFit Ki(apparent)

mM

R-7-fluorowarfarin Zhang et al., 1997 18 9.3 (20.3) 3.1 (20.8) 1.2 (21.2) 0.15 (22.0)S-7-fluorowarfarin Zhang et al., 1997 ,1 19 (1.3) 3.5 (0.5) 1.3 (0.1) 0.17 (20.8)Rac-6-fluorowarfarin Zhang et al., 1997 34 4.2 (20.9) 0.98 (21.5) 1.1 (21.5) 0.15 (22.3)Rac-6,7,8-trifluorowarfarin Zhang et al., 1997 5 7.9 (0.2) 1.4 (20.5) 4.6 (20.04) 0.19 (21.4)S-warfarin alcohol Zhang et al., 1997 7.5 4.7 (20.2) 3 (20.4) 2 (20.6) 0.16 (21.7)Ipriflavone Monostory et al., 1998 21.7 49 (0.3) 49 (0.3) 27 (0.1) 27 (0.1)7-hydroxyisoflavone Monostory et al., 1998 13.4 56 (0.6) 56 (0.6) 30 (0.3) 22 (0.2)Ticlopidine Donahue et al., 1997 38.8 49 (0.1) 33 (20.1) 17 (20.4) 16 (20.4)Omeprazole Ko et al., 1997 40.1 3.8 (21.0) 3.1 (21.1) 2.5 (21.2) 0.24 (22.2)Lansoprazole Ko et al., 1997 49.1 3.6 (21.1) 3.4 (21.1) 15 (20.5) 15 (20.5)Sertraline Schmider et al., 1997 33 8.3 (20.6) 4.8 (20.8) 16 (20.3) 16 (20.3)Desmethylsertraline Schmider et al., 1997 66 6.2 (21.0) 3.8 (21.2) 16 (20.6) 15 (20.6)Fluvoxamine Schmider et al., 1997 6 4.9 (20.1) 3.2 (20.3) 0.58 (21.0) 0.14 (21.6)Paroxetine Schmider et al., 1997 35 5.4 (20.8) 4.8 (20.9) 1.9 (21.3) 1.1 (21.5)

a Data used for model construction from Table 1.b Data used for model construction from Jones et al., 1996b.

TABLE 8

MS-WHIM PLS data for CYP2C9 data sets

MODEL CYP2C9a CYP2C9b CYP2C9c

Number of molecules in modeld 8/9 29/29 13/13Molecular surfacee weights 3,5 3,4,5,6 2,4LOOf q2 0.40 ,0.3 0.50LOO SDEPf 0.26 ,0.3 0.27LOO Confstat q2 0.54 0.64 0.45LOO Confstat SDEP 0.23 0.43 0.295RG Confstat (SD) ,0.3 0.55 (0.13) ,0.3Components 2 5 8

a Data used for model construction from Table 1.b Data used for model construction from Jones et al., 1996b.c Data used for model construction from Morsman et al., 1995.d Number of molecules used for this model out of the total number used in the original model.e Molecular surface weights: 15 Unitary, 25 positive potential, 35 negative potential, 45

H bond acceptor, 55 H bond donor, 65 hydrophobicity.f SDEP, standard deviation estimate predictions; Confstat, multiple conformers.

10013D- AND 4D-QSARS OF INHIBITORS FOR CYP2C9

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from

tures) within the CYP2C9 active site. This is in agreement with whatis known for substrates of CYP2C9 (Guengerich et al., 1986; Distel-rath and Guengerich, 1987; Poli-Scaife et al., 1997; Haining et al.,1999). Interestingly, these multiple substrate and inhibitor bindingdeterminants may explain the nonMichaelis-Menten kinetics dis-played for desmethylnaproxen (Tracy et al., 1997) andN-hydroxy-dapsone formation (Korzekwa et al., 1998). The ability of theseCYP2C9 substrates to illustrate both hetero- and homotropic activa-tion in vitro (Korzekwa et al., 1998) is similar to other CYPs (Ekinset al., 1998; Korzekwa et al., 1998). It is possible that the generationof a pharmacophore for substrates that display this non-Michaelis-Menten behavior alone may be necessary for the determination of thespecific structural features that elicit these kinetics, as previouslydemonstrated for CYP3A4 autoactivators (Ekins et al., 1999d).

The 3D- and 4D-QSAR modeling approaches taken in this studyrepresent an indirect examination of the CYP2C9 active site. Fromthese models we have been able to predict the interaction of somecompounds with this CYP. Additional studies will undoubtedly refinethese CYP2C9 models and may account for the compounds poorlypredicted. This will require an iterative approach of adding furtherdata to the model and evaluation with another set of test molecules. Inthe absence of a crystal structure for CYP2C9, researchers are limitedto computational models of substrates and inhibitors or active siteshomology-modeled on bacterial or other mammalian CYPs. Althoughthe multiple 3D- and 4D-QSAR approaches taken in this study werenot direct indicators of the active site, these results demonstrated thatstatistically predictive models of inhibitors of this CYP could begenerated. In addition, these models could provide further informationregarding the active site by the construction of a receptor surfacemodel. These procedures would potentially confirm the few structuralmodeling studies conducted with CYP2C9 as well as providing moreaccurate drug interaction predictions with this CYP active site. Thevalue of in silico approaches for predicting competitive type drug-drug interactions will become apparent after initial evaluation along-sideKi(apparent)values obtained with widely used in vitro systems.

Acknowledgments.We thank Dr. David Cummins for his help inthe implementation of the F-test in the PLS algorithm, Robert Conerfor assistance with permuting, and Dr. Patrick J. Murphy for his initialencouragement in pursuing this direction.

References

Baroni M, Costantino G, Cruciani G, Riganelli D, Valigi R and Clementi S (1993) Generatingoptimal linear PLS estimations (GOLPE): An advanced chemometric tool for handling3D-QSAR problems.Quant Struct-Act Rel12:9–20.

Bravi G, Gancia E, Mascani P, Pegna M, Todeschini R and Zaliani A (1997) MS-WHIM, new3D theoretical descriptors derived from molecular surface properties: A comparative 3DQSAR study in a series of steroids.J Comput-Aided Mol Des11:79–92.

Bravi G and Wikel JH (2000a) Application of MS-WHIM descriptors 1. Introduction of newmolecular surface properties. and 2. Prediction of binding affinity data.Quant Struct Act Rel19:29–38.

Bravi G and Wikel JH (2000b) Application of MS-WHIM Descriptors 3. Prediction of molecularproperties.Quant Struct Act Rel19:39–49.

Cramer RDI, DePriest SA, Patterson DE and Hecht P (1993) The developing practice of CoMFA,in 3D-QSAR in Drug Design: Theory, Methods and Applications, (Kubinyi H ed) pp 443–485,ESCOM, Leiden, the Netherlands.

Distelrath LM and Guengerich FP (1987) Enzymology of human cytochromes P-450, inMam-malian Cytochromes P-450, pp134–188, CRC Press, Boca Raton, FL.

Donahue SR, Flockhart DA, Abernethy DR and Ko J-W (1997) Ticlopidine inhibition ofphenytoin metabolism mediated by potent inhibition of CYP2C19.Clin Pharmacol Ther62:572–577.

Ekins S, Bravi G, Binkley S, Gillespie JS, Ring BJ, Wikel JH and Wrighton SA (1999a) Threeand four dimensional-quantitative structure activity relationship (3D/4D-QSAR) analyses ofCYP2D6 inhibitors.Pharmacogenetics9:477–489.

Ekins S, Bravi G, Binkley S, Gillespie JS, Ring BJ, Wikel JH and Wrighton SA (1999b) Three-and four- dimensional-quantitative structure activity relationship analyses of CYP3A4 inhib-itors. J Pharmacol Exp Ther290:429–438.

Ekins S, Bravi G, Ring BJ, Gillespie TA, Gillespie JS, VandenBranden M, Wrighton SA andWikel JH (1999c) Three dimensional-quantitative structure activity relationship (3D-QSAR)analyses of substrates for CYP2B6.J Pharmacol Exp Ther288:21–29.

Ekins S, Bravi G, Wikel JH and Wrighton SA (1999d) Three-dimensional-quantitative structureactivity relationship analyses of CYP3A4 substrates.J Pharmacol Exp Ther291:424–433.

Ekins S, Ring BJ, Binkley SN, Hall SD and Wrighton SA (1998) Autoactivation and activationof the cytochrome P450s.Int J Clin Pharmacol Ther36:642–651.

Guengerich FP, Shimada T, Umbenhauer DR, Martin MV, Misono KS, Distelrath LM, ReillyPEB and Wolff T (1986) Structure and function of cytochrome P-450.Adv Exp Med Biol197:83–94.

Haining RL, Jones JP, Henne KR, Fisher MB, Koop DR, Trager WF and Rettie AE (1999)Enzymatic determinants of the substrate specificity of CYP2C9: Role of B’-C loop residues inproviding thep-stacking anchor site for warfarin binding.Biochemistry38:3285–3292.

Hall SD, Hamman MA, Rettie AE, Wienkers LC, Trager WF, VandenBranden M and WrightonSA (1994) Relationships between the levels of cytochrome P4502C9 and its prototypiccatalytic activities in human liver microsomes.Drug Metab Dispos22:975–977.

Jones BC, Hawksworth G, Horne V, Newlands A, Tute M and Smith DA (1993) Putative activesite model for CYP2C9 (tolbutamide hydroxylase).Br J Clin Pharmacol34:143–144P.

Jones BC, Hawksworth G, Horne VA, Newlands A, Morsman J, Tute MS and Smith DA (1996a)Putative active site template model for cytochrome P4502C9 (tolbutamide hydroxylase).DrugMetab Dispos24:260–266.

Jones JP, He M, Trager WF and Rettie AE (1996b) Three-dimensional quantitative structure-activity relationship for inhibitors of cytochrome P4502C9.Drug Metab Dispos24:1–6.

Klein CDP and Hopfinger AJ (1998) Pharmacological activity and membrane interactions ofantiarrhythmics: 4D-QSAR/QSPR analysis.Pharm Res15:303–311.

Klose TS, Ibeanu GC, Ghanayem BI, Pedersen LG, Li L, Hall SD and Goldstein JA (1998)Identification of residues 286 and 289 as critical for conferring substrate specificity of humanCYP2C9 for diclofenac and ibuprofen.Arch Biochem Biophys357:240–248.

Ko J-W, Sukhova N, Thacker D, Chen P and Flockhart DA (1997) Evaluation of omeprazole andlansoprazole as inhibitors of cytochrome P450 isoforms.Drug Metab Dispos25:853–862.

Korzekwa KR, Krishnamachary N, Shou M, Ogai A, Parise RA, Rettie AE, Gonzalez FJ andTracy TS (1998) Evaluation of atypical cytochrome P450 kinetics with two-substrate models:Evidence that multiple substrates can simultaneously bind to cytochrome P450 active sites.Biochemistry37:4137–4147.

Kubinyi H (1997) QSAR and 3D QSAR in drug design Part 2: Applications and problems.DrugDiscov Today2:538–546.

Leeman T, Transon C and Dayer P (1992) Cytochrome P450TB (CYP2C): A major monooxy-genase catalyzing diclofenac 49-hydroxylation in human liver.Life Sci52:29–34.

Lewis DFV, Dickins M, Weaver RJ, Eddershaw PJ, Goldfarb PS and Tarbit MH (1998)Molecular modelling of human CYP2C subfamily enzymes CYP2C9 and CYP2C19: Ratio-nalization of substrate specificity and site-directed mutagenesis experiments in the CYP2Csubfamily.Xenobiotica28:235–268.

Mancy A, Broto P, Dijols S, Dansette PM and Mansuy D (1995) The substrate binding site ofhuman liver cytochrome P4502C9: An approach using designed tienilic acid derivatives andmolecular modeling.Biochemistry34:10365–10375.

Miners JO and Birkett DJ (1998) Cytochrome P4502C9: An enzyme of major importance inhuman drug metabolism.Br J Clin Pharmacol45:525–538.

Miners JO, Smith KJ, Robson RA, McManus ME, Veronese ME and Birkett DJ (1988)Tolbutamide hydroxylation by human liver microsomes. Kinetic characterization and relation-ship to other cytochrome P-450 dependent xenobiotic oxidations.Biochem Pharmacol37:1137–1144.

Monostory K, Vereczkey L, Levai F and Szatmari I (1998) Ipriflavone as an inhibitor of humancytochrome P450 enzymes.Br J Pharmacol123:605–610.

Morsman JM, Smith DA, Jones BC and Hawksworth GM (1995) Role of hydrogen-bonding insubstrate structure-activity relationships for CYP2C9.ISSX Proc8:259.

Newlands AJ, Smith DA, Jones BC and Hawksworth GM (1992) Metabolism of non-steroidalanti-inflammatory drugs by cytochrome P450 2C.Br J Clin Pharmacol34:152P.

Poli-Scaife S, Attias R, Dansette PM and Mansuy D (1997) The substrate binding site of humanliver cytochrome P4502C9: An NMR study.Biochemistry36:12672–12682.

Quintana J, Contijoch M, Cuberes R and Frigola J (1995) Structure-activity relationships andmolecuar modelling studies of a series of H1 antihistamines, inQSAR and MolecularModelling: Concepts, Computational Tools and Biological Applications. pp 282–288, ProusScience Publishers, Barcelona, Spain.

Ring BJ, Binkley SN, Roskos L and Wrighton SA (1995) Effect of fluoxetine, norfluoxetine,sertraline and desmethyl sertraline on on human CYP3A catalyzed 19-hydroxy midazolam.J Pharmacol Exp Ther275:1131–1135.

Ring BJ, Binkley SN, Vandenbranden M and Wrighton SA (1996) In vitro interaction of theantipsychotic agent olanzapine with human cytochromes P450 CYP2C9, CYP2C19, CYP2D6and CYP3A.Br J Clin Pharmacol41:181–186.

Schmider J, Greenblatt DJ, von Moltke LL, Karsov D and Shader RI (1997) Inhibition ofCYP2C9 by selective serotonin reuptake inhibitors in vitro: Studies of phenytoin p-hydroxylation.Br J Clin Pharmacol44:495–498.

Smith DA, Ackland MJ and Jones BC (1997a) Properties of cytochrome P450 isoenzymes andtheir substrates. Part 1: Active site characteristics.Drug Discov Today2:406–414.

Smith DA, Ackland MJ and Jones BC (1997b) Properties of cytochrome P450 isoenzymes andtheir substrates. Part 2: Properties of cytochrome P450 substrates.Drug Discov Today2:479–486.

Smith DA and Jones BC (1992) Speculations on the substrate structure-activity relationship(SSAR) of cytochrome P450 enzymes.Biochem Pharmacol44:2089–2098.

Tracy TS, Marra C, Wrighton SA, Gonzalez FJ and Korzekwa KR (1997) Involvement ofmultiple cytochrome P450 isoforms in naproxen O-demethylation.Eur J Clin Pharmacol52:293–298.

van der Hoeven TA and Coon M (1974) Preparation and properties of partially purifiedcytochrome P-450 and reduced nicotinamide adenine dinucleotide phosphate-cytochromeP450-reductase from rabbit liver microsomes.J Biol Chem249:6302–6310.

Weininger D (1988) SMILES 1. Introduction and encoding rules.J Chem Inf Comput Sci28:31.Wold S, Johansson E and Cocchi M (1993) PLS-Partial least squares projections to latent

structures, in3D-QSAR in Drug Design: Theory, Methods and Applications(Kubinyi H ed) pp523–550, ESCOM, Leiden, the Netherlands.

Zhang ZY, Kerr J, Wexler RS, Li H-Y, Robinson AJ, Harlow PP and Kaminsky LS (1997)Warfarin analog inhibition of human CYP2C9-catalyzed S-warfarin 7-hydroxylation.ThrombRes88:389–398.

1002 EKINS ET AL.

at ASPE

T Journals on June 14, 2018

dmd.aspetjournals.org

Dow

nloaded from