Downloaded from
`
`dmd.aspetjournals.org
`
` at ASPET Journals on April 14, 2020
`
`0090-9556/02/3001-7–12$3.00
`DRUG METABOLISM AND DISPOSITION
`Copyright © 2002 by The American Society for Pharmacology and Experimental Therapeutics
`DMD 30:7–12, 2002
`
`Vol. 30, No. 1
`564/951813
`Printed in U.S.A.
`
`COMPUTATIONAL MODELS FOR CYTOCHROME P450: A PREDICTIVE ELECTRONIC
`MODEL FOR AROMATIC OXIDATION AND HYDROGEN ATOM ABSTRACTION
`
`JEFFREY P. JONES, MICHAEL MYSINGER, AND KENNETH RAY KORZEKWA
`
`Department of Chemistry, Washington State University, Pullman, Washington (J.P.J.); and Camitro Corporation,
`Menlo Park, California (M.M., K.R.K.)
`
`(Received August 20, 2001; accepted October 1, 2001)
`
`This paper is available online at http://dmd.aspetjournals.org
`
`ABSTRACT:
`
`Experimental observations suggest that electronic characteristics
`play a role in the rates of substrate oxidation for cytochrome P450
`enzymes. For example, the tendency for oxidation of a certain
`functional group generally follows the relative stability of the rad-
`icals that are formed (e.g., N-dealkylation > O-dealkylation > 2°
`carbon oxidation > 1° carbon oxidation). In addition, results show
`that useful correlations between the rates of product formation
`can be developed using electronic models. In this article, we at-
`tempt to determine whether a combined computational model for
`
`aromatic and aliphatic hydroxylation can be developed. Toward
`this goal, we used a combination of experimental data and
`semiempirical molecular orbital calculations to predicted activa-
`tion energies for aromatic and aliphatic hydroxylation. The result-
`ing model extends the predictive capacity of our previous aliphatic
`hydroxylation model to include the second most important group
`of oxidations, aromatic hydroxylation. The combined model can ac-
`count for about 83% of the variance in the data for the 20 compounds
`in the training set and has an error of about 0.7 kcal/mol.
`
`The P4501 enzymes are a superfamily of monooxygenases involved
`in the metabolism of both exogenous and endogenous compounds.
`Ironically, these enzymes play a central role in both the prevention
`and induction of chemical toxicities and carcinogenicity. Although
`most P450 oxidations of xenobiotics result in detoxification, occa-
`sionally a more toxic intermediate is formed. In fact, many ultimate
`toxins and carcinogens are formed by the bioactivation of less reactive
`compounds, and many bioactivation reactions are mediated by the
`P450 enzymes. Often bioactivation reactions are in competition with
`detoxification pathways for the same substrate. Since these enzymes
`play such a central role in both detoxification and bioactivation,
`predictive models for cytochrome P450 catalysis will be useful tools
`for evaluating of the potential risks of environmental exposures. One
`of the most pertinent but difficult problems in risk assessment is
`translating bench results and mechanistic information into a form that
`can be used. This article outlines steps toward the development of
`computational models from laboratory data into a form that can be
`used for risk assessments that accurately reflect experimental results.
`For example, these models now more completely describe all posi-
`tions of metabolism for nitriles and should be more complete in
`predicting the toxicity related to nitrile metabolism. These semiem-
`pirical computational models blend experimental data and computa-
`tional chemistry in such a way as to provide a consistent prediction of
`the bioactivation rates for a broad spectrum of compounds. In partic-
`ular, these models can be used to predict xenobiotic metabolism by the
`
`This work was supported by National Institute of Environmental Health Sci-
`ences Grant ES09122.
`1 Abbreviations used are: P450, cytochrome P450; AM1, Austin model-1; PNR,
`p-nitrosophenoxy radical.
`
`Address correspondence to: Jeffrey P. Jones, Department of Chemistry,
`Washington State University, Pullman, Washington, 99164. E-mail: jpj@wsu.edu
`
`P450 enzyme family, including the bioactivation of compounds to
`toxins and carcinogens.
`Models such as those presented here for P450-mediated reac-
`tions can also play a role in drug design. Tools that predict
`regioselectivity can be used to assess the pharmacokinetics of
`drugs before synthesis, saving time and money in the drug devel-
`opment process. These types of tools, either alone or in combina-
`tion with homology or pharmacophore models, can also provide for
`the design of better drugs, with higher compliance and fewer toxic
`side-effects. However, P450 enzymes are difficult to model by the
`standard methods used for most drug targets, which are based
`mainly on predicting binding affinities related to steric features,
`since accurate results depend upon the prediction of both the
`electronic and steric features of the enzymes. This is different from
`many other enzyme systems since P450 has the need to metabolize
`a vast array of xenobiotics, which makes it impractical to have one
`enzyme for each compound or even each class of compounds.
`Therefore, although most cellular functions tend to be very spe-
`cific, xenobiotic metabolism requires enzymes with diverse sub-
`strate specificities. The cytochromes P450 have apparently as-
`sumed much of this role. The required diversity is accomplished by
`families and subfamilies of enzymes with generally broad substrate
`specificities, a very reactive oxygenating species and a broad
`regioselectivity. Thus, for many reactions, the electronic features
`of the substrate are all that are required to predict regioselectivity
`(Grogan et al., 1992; Harris et al., 1992; de Groot et al., 1995,
`1999; Yin et al., 1995).
`We have reported a rapid electronic model for the prediction of
`regioselectivity in P450-mediated hydrogen atom abstraction mecha-
`nisms (Korzekwa et al., 1990a). The methods depend only on the
`calculation of the AM1 ground-state energies for the parent compound
`
`7
`
`Apotex Ex. 1016
`
`Apotex v. Auspex
`IPR2021-01507
`
`

`

`Downloaded from
`
`dmd.aspetjournals.org
`
` at ASPET Journals on April 14, 2020
`
`8
`
`JONES ET AL.
`
`TABLE 1
`
`AM1 energies for addition of methoxy radical to aromatic compounds
`
`Compound
`
`Benzene
`Toluene (para)
`Toluene (meta)
`Toluene (ortho)
`Anisole (meta)
`Anisole (para)
`Anisole (ortho)
`Chlorobenzene (para)
`Chlorobenzene (meta)
`Aniline (para)
`Aniline (meta)
`Nitrobenzene (para)
`Nitrobezene (meta)
`Cyanobenzene (para)
`Cyanobenzene (meta)
`Cyanobenzene (ortho)
`o-Xylene (para)
`2-Methylanisole (4-position)
`Ethylbenzene (para)
`p-Xylene (ortho)
`Napthalene (ortho)
`Napthalene (para)
`Benzimidazole (ortho)
`Benzimidazole (meta)
`Benzimidazole (para)
`
`a
`Hgrd
`
`0.035
`0.023
`0.023
`0.023
`⫺0.025
`⫺0.025
`⫺0.025
`0.023
`0.023
`0.032
`0.032
`0.040
`0.040
`0.085
`0.085
`0.085
`0.012
`⫺0.036
`0.013
`0.011
`0.064
`0.064
`0.107
`0.107
`0.107
`
`b
`
`Htran
`
`0.036
`0.023
`0.023
`0.023
`⫺0.025
`⫺0.026
`⫺0.028
`0.023
`0.024
`0.030
`0.032
`0.042
`0.042
`0.086
`0.086
`0.087
`0.011
`⫺0.037
`0.013
`0.010
`0.058
`0.059
`0.104
`0.106
`0.106
`
`a The ground state energy of the parent compound in Hartrees/mol.
`b The transition state energy for methoxy addition in Hartrees/mol.
`c The heat of formation of the tetrahedral intermediate after methoxy addition in Hartrees/mol.
`d The enthalpy of activation in kcal/mol.
`e The enthalpy of reaction in kcal/mol.
`
`c
`
`Htetrs
`⫺0.012
`⫺0.025
`⫺0.024
`⫺0.024
`⫺0.071
`⫺0.075
`⫺0.075
`⫺0.025
`⫺0.023
`⫺0.019
`⫺0.014
`⫺0.006
`⫺0.005
`0.037
`0.039
`0.039
`⫺0.036
`⫺0.085
`⫺0.034
`⫺0.037
`0.006
`0.010
`0.053
`0.057
`0.059
`
`d
`Hact
`
`9.4
`9.0
`9.3
`9.1
`9.5
`8.5
`7.8
`9.0
`9.4
`7.9
`9.2
`10.2
`10.5
`9.5
`9.8
`10.3
`8.9
`8.8
`9.0
`8.9
`5.3
`5.8
`7.3
`8.3
`8.7
`
`e
`Hreac
`⫺20.3
`⫺21.2
`⫺20.3
`⫺20.5
`⫺19.8
`⫺21.8
`⫺21.7
`⫺21.4
`⫺20.4
`⫺22.9
`⫺20.0
`⫺20.1
`⫺19.1
`⫺20.9
`⫺19.8
`⫺19.6
`⫺21.1
`⫺21.4
`⫺21.0
`⫺20.6
`⫺27.6
`⫺25.0
`⫺24.6
`⫺22.1
`⫺20.8
`
`and the product radicals. The relative activation energy is predicted by
`the use of eq. 1.
`
`⌬Hact ⫽ 2.60 ⫹ 0.22共⌬H兲R ⫹ 2.38共IPrad)
`
`(1)
`
`This model has been successfully used to predict the rates of metab-
`olism of halogenated hydrocarbons (Harris et al., 1992; Yin et al.,
`1995) and the regioselectivity of nitrile metabolism (Grogan et al.,
`1992) both in vivo and in vitro. In this article, we report the devel-
`opment of a model for aromatic oxidation and the combination of this
`model with our previous model for hydrogen atom abstraction. The
`combined models allow for the prediction of the electronic component
`for the regioselectivity of most of the reactions catalyzed by the
`cytochrome P450 superfamily of enzymes. These models are general
`models for the electronic component of P450 enzymes, which we have
`shown to be rather constant across both the enzyme families and
`species (Jones et al., 1990; Karki et al., 1995; Yin et al., 1995; Higgins
`et al., 1998).
`
`Materials and Methods
`
`Data analysis and graphing were done with SigmaPlot version 5 (SPSS, Inc.
`Chicago, IL). The energy difference for regioselectivity was determined by
`taking the log of the ratio of metabolites and multiplying by 0.616, which gives
`the energetic difference at 310°K.
`Semiempirical calculations were performed with Gaussian 98 (Gaussian,
`Inc., Pittsburgh, PA.). The semiempirical activation energies for hydrogen
`atom abstraction for the reactions outlined in the text were estimated with our
`quantum mechanical model [the p-nitrosophenoxy radical (PNR) model] based
`on hydrogen abstraction reactions using PNR (Korzekwa et al., 1990b) and the
`semiempirical AM1 Hamiltonian (Dewar et al., 1985). The model uses the
`calculated (AM1) enthalpies of reaction and the ionization potentials of the
`resultant radicals to predict the AM1 activation enthalpies (Hact). All open shell
`systems were treated with an unrestricted Hartree-Fock Hamiltonian. Each of
`the reactants and the transition states for the aromatic hydroxylation reactions
`were minimized using the default procedures in Gaussian 98, and each tran-
`
`sition state was confirmed by frequency calculations and had only one negative
`eigenvalue. All transition states were found by following the most negative
`eigenvalue after the methoxy radical oxygen-aromatic carbon bond in the
`tetrahedral intermediate was lengthened to 1.95 Å.
`The aromatics compounds listed in Table 1 were selected to cover a range
`of electron-donating and electron-withdrawing groups based on Hammett ␴
`values. We emphasized single-ring compounds at the expense of multiple-ring
`compounds since all of the experimental data were for single-ring-containing
`compounds. It is possible that a different model will be required for multiple-
`ring-containing compounds, five-membered rings, and other types of aromatic
`compounds for which no experimental is available.
`
`Results
`Ideally, we would be able to use the same oxidant (PNR) to model
`P450 aromatic addition that we used for hydrogen atom abstraction.
`We have shown previously that aromatic oxidation is likely to occur
`by addition of a triplet-like oxygen to form an tetrahedral intermediate
`(Korzekwa et al., 1989). This tetrahedral intermediate can then rear-
`range to epoxides, ketones, and phenols. Although addition of the
`PNR to an aromatic carbon progresses smoothly through a transition
`state to a tetrahedral intermediate, the electronic state of the tetrahe-
`dral intermediate is not the ground state. Thus, the reactions from
`reactants to products and products to reactants occur on two different
`potential energy surfaces. In addition to complicating any transition
`state studies, the value of any Brønsted correlations will be in doubt
`since the transition states are not necessarily intermediate between
`products and reactants. Thus, the p-nitrosophenoxy radical alone is
`not likely to serve as a versatile model for other P450-mediated
`oxidations, and other small oxygen radicals were tested within the
`AM1 formalism.
`To circumvent the problems associated with PNR additions to
`aromatic systems, methoxy radical was used for aromatic addition
`reactions. Using the AM1 formalism,
`the methoxy radical adds
`smoothly to aromatic compounds and, in contrast to the p-nitrosophe-
`
`Apotex Ex. 1016
`
`

`

`COMPUTATIONAL MODELS FOR CYTOCHROME P450
`
`9
`
`Downloaded from
`
`dmd.aspetjournals.org
`
` at ASPET Journals on April 14, 2020
`
`been shown to undergo rapid interchange between the two sites of
`interest by isotope effect experiments. The octane and hexane values
`are for C-1 versus C-2 hydroxylation, whereas ethylbenzene and
`diphenylpropane provide an estimate of the effect of a phenyl sub-
`stituent on C-1 versus C-2 hydroxylation and benzylic versus an
`internal secondary position. The remaining compounds give values for
`substituent effect on benzylic hydroxylation and aromatic O-dealky-
`lation. The computationally predicted energies are given in Table 2. A
`plot of the predicted versus the observed regioselectivity is given in
`Fig. 2. We performed a linear regression on the experimental versus
`theoretical data to correct for the differences in predicted versus
`theoretical regioselectivities, forcing the y-intercept through the ori-
`gin. The results are shown in eq. 3.
`
`⌬⌬G measured ⫽ 0.74 䡠 共⌬Hact(Habs1) ⫺ ⌬Hact(Habs2))
`
`(3)
`
`Next, we obtained from the literature different experimentally mea-
`sured regioselectivity reactions that involved hydrogen atom abstrac-
`tion and aromatic oxidation. This data are shown in Table 3. The
`methods used for determining the energies have been described pre-
`viously (Higgins et al., 2001). Ideally, the regioselectivity of reaction
`is measured in multiple enzymes known to show relatively nonspe-
`cific binding. The ratio of two metabolites then is assumed to be a
`measure of the difference in reactivity. In practice, only a very small
`number of reported regioselectivity values meet these criteria. The
`sparse dataset means that as more data becomes available the models
`are likely to change; however, the approach described here should still
`be applicable.
`The metabolism of toluene was determined to give around 70%
`benzyl alcohol and 30% aromatic hydroxylation in phenobarbital-
`induced rat microsomes (Hanzlik et al., 1984; Hanzlik and Ling,
`1990) and to give 95% benzyl alcohol and 5% aromatic hydroxylation
`in human microsomes (Tassaneeyakul et al., 1996). These results give
`differences in energy of activation for the aromatic to benzylic hy-
`droxylation of 0.7 and 1.8 kcal/mol for the two systems, respectively.
`Since the work done in rat microsomes gave a range of ortho, meta,
`and para products, we decided to take the mean of the upper and lower
`limit of the range to fit to the computational results. For ethylbenzene,
`the reported ratio of para hydroxylation to benzylic hydroxylation in
`purified CYP2B1 was around 0.0013 with an energy difference of 4.1
`kcal/mol, reflecting the faster reaction for the secondary benzylic
`position relative to the primary benzylic position of toluene (White et
`al., 1986). For o-xylene and p-xylene the ratios of benzylic to aromatic
`were measured to be 11.5 and 33, which correspond to energy differ-
`ences of 1.5 and 2.2 kcal/mol (Iyer et al., 1997). When a similar
`experiment was conducted by Lindsey-Smith for anisole metabolism,
`aromatic oxidation was found to be faster than O-dealkylation by
`about 3 times (Lindsay-Smith and Sleath, 1983), corresponding to a
`difference in energy of 0.67 kcal/mol. This is reduced to a ratio of
`around 2 for ortho hydroxylation of 2-methylanisole, when the aver-
`age value of single-expressed enzymes shown to give information
`about energetics is taken into account (Higgins et al., 2001). Of these
`experiments, only the work of Higgins et al. (2001) was conducted
`with the expressed purpose of determining the energy difference
`between two pathways, although ideally all of the data would be
`generated in the same enzyme preparation and with the goal of
`accurately determining the ratios of interest.
`With these data, the model for hydrogen atom abstraction and the
`new model for aromatic addition can be combined. Figure 3 is a plot
`of the predicted versus measured difference in energy for the oxida-
`tions in Table 3. Obviously, a linear relationship exists. The equation
`for the line is shown in eq. 4. In eq. 3, ⌬⌬Gmeasured is the measured
`
`FIG. 1.The linear correlation of aromatic activation energies and the heats of
`reaction for formation of a tetrahedral intermediate by methoxy radical.
`The R2 value is 0.92, and the 95% confidence limits are shown by the dashed
`lines.
`
`noxy radical, remains on a single potential energy surface. Methoxy
`radical was added to the aromatic compounds shown in Table 1, and
`the heats of reaction and transition state were characterized for each
`reaction. A good linear correlation was observed for the heats of
`reaction and the activation energies reported in Table 1. A plot of the
`data is shown in Fig. 1. Equation 2 is the equation that describes the
`linear correlation between activation energy and heats of reaction.
`This linear correlation has a R2 value of 0.92 and a standard error
`approximately 0.35 kcal/mol. In eq. 2, ⌬Hact is the enthalpy of
`activation, and ⌬Hreac is the enthalpy of reaction. Since this equation
`depends only on the enthalpy of reaction for predicting rate, even the
`activation energy of a large compound can be predicted extremely
`rapidly.
`
`⌬Hact ⫽ 21.91 ⫹ 0.61 䡠 ⌬Hreac
`
`(2)
`
`Thus, correlations between computational activation energies and
`computational heats of reaction can be obtained for both hydrogen
`abstraction reactions (eq. 1) and aromatic addition reactions (eq. 2)
`with separate regression models. However, the presence of Brønsted
`correlations for two different series of reactions does not necessarily
`imply that the activation energies can be directly compared. For
`example, the hydrogen atom abstractions are predicted by AM1 to
`have barriers around 20 kcal/mol, whereas aromatic oxidation has
`barriers around 10 kcal/mol. Since experimentally hydrogen atom
`abstraction is found to be favored over aromatic hydroxylation in
`many instances, these results are quantitatively erroneous. Therefore,
`a method is required to parameterize the two models to a common
`energy scale.
`The activation energies of our aliphatic model obtained from eq. 1
`has never been corrected to reflect experimental activation energies.
`Thus, we need to scale eq. 1 relative to experimentally measured
`values. Literature values for regioselectivity that appear to reflect
`intrinsic reactivity differences were found for octane (Jones et al.,
`1986, 1990), hexane (Morohashi et al., 1983), diphenylpropane
`(Hjelmeland et al., 1977a), substituted diphenylpropanes (Hjelmeland
`et al., 1977b), 2-methylanisole (Higgins et al., 2001), 4-methyanisole
`(Higgins et al., 2001), ␣-chloro-p-xylene (Higgins et al., 2001), and
`ethylbenzene (White et al., 1986). The values for octane, 2-methyl-
`anisole, 4-methyanisole, ␣-chloro-p-xylene, and ethylbenzene have
`
`Apotex Ex. 1016
`
`

`

`10
`
`JONES ET AL.
`
`TABLE 2
`
`Experimentally measured energy differences in activation energies for aliphatic P450 oxidation reactions
`
`Compound
`
`Reference
`
`Regioselectivity Positions
`
`Ethylbenzene
`␣-Chloro-p-xylene
`2-Methylanisole
`4-Methylanisole
`1,3-Diphenylpropane
`Hexane
`Octane
`1-Phenyl-3-(4-fluorophenyl)propane
`1-Phenyl-3-(4-methylphenyl)propane
`1-Phenyl-3-(4-trifluoromethylphenyl)propane
`
`White et al., 1996
`Higgins et al., 2001
`Higgins et al., 2001
`Higgins et al., 2001
`Hjelmeland et al., 1977
`Morohashi et al., 1983
`Jones et al., 1990
`Hjelmeland et al., 1977
`Hjelmeland et al., 1977
`Hjelmeland et al., 1977
`
`Primary/benzylic
`Benzylic/CI-methyl benzylic
`Benzylic/O-demethylation
`Benzylic/O-demethylation
`Benzylic/secondary
`1/2 Hexanol
`1/2 Octanol
`Benzylic/substitutedc
`Benzylic/substitutedc
`Benzylic/substitutedc
`
`Measured ⌬
`kcal/mola
`⫺4.39
`⫺0.74
`0.40
`0.72
`1.28
`⫺1.85
`⫺2.14
`0.51
`0.0
`1.81
`
`Predicted ⌬
`kcal/molb
`⫺4.58
`⫺1.31
`2.25
`2.21
`2.44
`⫺2.15
`⫺2.14
`⫺0.16
`⫺0.34
`0.82
`
`a The energy difference as determined by taking the ln of the measured ratios and multiplying by 0.616, as described under Materials and Methods.
`b The energy difference between the predicted activation energy for hydrogen atom abstractions based on Eq. 2.
`c Benzylic hydroxylation of the phenyl ring over benzylic hydroxylation of the substituted phenyl ring.
`
`Downloaded from
`
`dmd.aspetjournals.org
`
` at ASPET Journals on April 14, 2020
`
`1998). This apparent contradiction can be resolved by an understand-
`ing of the kinetics of branched pathways, which was described by
`Jones et al. (1986). In brief, if product formation is not rate limiting,
`the reactivity of the substrate will not affect the rate of product
`formation. However, if an alternate product can be formed, differ-
`ences in reactivity will be observed. For example, if norcamphor is
`substituted with deuterium, less product is observed than if it has
`hydrogen, even though the rate-limiting step is not norcamphor hy-
`droxylation. Atkins and Sligar (1987) found that the isotope effect was
`observed because the build-up in the enzyme-substrate complex due
`to deuterium substitution was prevented by alternate product forma-
`tion. In this case, the alternate product was water. We have seen
`similar results for the C-1 hydroxylation of octane; however in this
`case, the alternate products were C-2 and C-3 oxidation products
`(Jones et al., 1986). Therefore, slowing the rate of oxidation of one
`position by deuterium substitution causes an increase in the rate of
`metabolism of another position or an increase in decoupling to water
`formation (Higgins et al., 1998). This can only occur if the rate of
`exchange of the different positions within the active site is as fast or
`faster than the rate of the oxidation step. This means that for a
`substrate that can rotate freely in the active site the regioselectivity
`will be primarily determined by the electronic reactivity of the various
`positions on the molecule (Higgins et al., 2001). For these molecules,
`product regioselectivities can be predicted by the electronic charac-
`teristics of the substrate. Conversely, regioselectivity can provide an
`important tool for testing computational models.
`We previously reported the development of an electronic model for
`hydrogen atom abstraction with can predict aliphatic hydroxylation,
`amine dealkylation, and O-dealkylation (Korzekwa et al., 1990a). To
`our knowledge, this article is the first time anyone has combined
`predictive models for P450-mediated aromatic and aliphatic oxidation
`so that a continuous prediction can be made of either aromatic,
`aliphatic, or a combination of these mechanisms (Fig. 4). However,
`qualitative models that approach these models have been described for
`CYP2D6 (de Groot et al., 1999). The model described here was
`generated by correcting the computational predictions with experi-
`mental results for microsomes and single expressed enzymes. Every
`effort was made to use only experimental data that would be expected
`to be free of specific enzyme orientation effects, as described by
`Higgins et al. (2001), and thus should be a general electronic model
`for cytochrome P450 enzymes. However, we did use a number of data
`points derived from microsomal preparations. It is possible that the
`two metabolites in these cases result from two distinct enzymes in the
`microsomal preparation, and when possible, we chose the phenobar-
`bital-induced microsomal preparation since the major isoform is
`CYP2B1, an enzyme that has shown a lack of binding preference and
`
`FIG. 2.A plot of predicted and measured energy differences for hydrogen atom
`abstraction from the values given in Table 2.
`The R2 value is 0.81, and the 95% confidence limits are shown by the dashed
`lines.
`
`difference in energies based on regioselectivity, ⌬Hact(arom) is the
`predicted activation energy for aromatic hydroxylation based from eq.
`2, and ⌬Hact(Habs) is the uncorrected AM1 activation energy for
`hydrogen atom abstraction. The fit eq. 3 gives an R2 value of 0.85,
`with a standard error approximately 0.65 kcal/mol.
`
`⌬⌬G measured ⫽ 1.22 䡠 共⌬Hact(Habs)兲 ⫺ 1.1 䡠 共⌬Hact(arom)兲 ⫺ 17.5
`
`(4)
`
`Finally, predicted regioselectivity for both eq. 3 and eq. 4 can be
`plotted together as shown in Fig. 4. The R2 value for the data is 0.83,
`with a standard error approximately 0.71 kcal/mol.
`
`Discussion
`Experimental observations suggest that electronic characteristics
`play a role in the rates of substrate oxidation for cytochrome P450
`enzymes. For example, the tendency for oxidation of a certain func-
`tional group generally follows the relative stability of the radicals that
`are formed (e.g., N-dealkylation ⬎ O-dealkylation ⬎ 2° carbon oxi-
`dation ⬎ 1° carbon oxidation). In addition, results show that useful
`correlations between the rates of product formation can be developed
`using electronic models (Grogan et al., 1992; Yin et al., 1995). These
`results are apparently contradicted by the results of kinetic experi-
`ments, which show the steps before substrate oxidation (electron
`reduction) to be rate limiting in the catalytic cycle (Higgins et al.,
`
`Apotex Ex. 1016
`
`

`

`COMPUTATIONAL MODELS FOR CYTOCHROME P450
`
`11
`
`Experimentally measured energy differences in activation energies for aromatic and aliphatic P450 oxidation reactions
`
`TABLE 3
`
`Compound
`
`Ethylbenzene
`Toluene
`p-Xylene
`o-Xylene
`Anisole
`Toluene
`Toluene
`2-Methylanisole
`Anisole
`Anisole
`
`Reference
`
`White et al., 1986
`Hanzlik et al., 1984
`Iyer et al., 1997
`Iyer et al., 1997
`Hanzlik et al., 1984
`Hanzlik et al., 1984
`Hanzlik et al., 1984
`Higgins et al., 2001
`Hanzlik et al., 1984
`Hanzlik et al., 1984
`
`Regioselectivity Positions
`
`Aromatic (para)/benzylic
`Aromatic (meta)/benzylic
`Aromatic (ortho)/benzylic
`Aromatic (para)/benzylic
`Aromatic (meta)/O-demethylation
`Aromatic (ortho)/benzylic
`Aromatic (para)/benzylic
`Aromatic (para)/O-demethylation
`Aromatic (ortho)/O-demethylation
`Aromatic (para)/O-demethylation
`
`Measureda
`⫺4.1
`⫺2.3
`⫺2.2
`⫺1.5
`⫺0.9
`⫺0.9
`⫺0.6
`0.17
`0.25
`1.02
`
`Predictedb
`⫺3.2
`⫺1.9
`⫺2.3
`⫺1.8
`⫺0.6
`⫺1.7
`⫺1.4
`0.5
`0.7
`0.8
`
`a The energy difference in kcal/mol as determined by taking the ln of the measured ratios and multiplying by 0.616, as described under Materials and Methods.
`b The energy difference between the predicted activation energy for aromatic addition based on Eq. 3 and the predicted activation energy for hydrogen atom abstractions based on Eq. 2.
`
`Downloaded from
`
`dmd.aspetjournals.org
`
` at ASPET Journals on April 14, 2020
`
`FIG. 3.A plot of the energy predicted differences between aliphatic and
`aromatic positions versus measured difference in energy for the oxidations in
`Table 3.
`The R2 value is 0.85, and the 95% confidence limits are shown by the dashed
`lines.
`
`FIG. 4.A plot of the predicted versus actual values for the combined aromatic
`and hydrogen atom abstraction model.
`The R2 value is 0.83, and the 95% confidence limits are shown by the dashed
`lines.
`
`a dominance of electronic factors for the type of compounds use in
`this study (White et al., 1984; Higgins et al., 2001). However, to
`exclude this data would decrease an already small dataset. In general,
`we have attempted to construct our electronic models with data for
`compounds that are small and do not contain strong binding features
`(Higgins et al., 2001). This means that the best data will not have
`orienting effects associated with the enzyme tertiary structure. Ori-
`enting effects have been shown to be important for a number of P450
`enzymes and need to be added to the electronic models for accurate
`prediction of metabolism by these enzymes (Szklarz and Halpert,
`1997; de Groot et al., 1999; Lightfoot et al., 2000; Rao et al., 2000;
`Szklarz et al., 2000; Xue et al., 2001).
`Thus, although this model is general to the P450 superfamily, it is
`not going to work to any significant extent on enzymes such as those
`in the 4B family that have a large contribution from substrate-protein
`interactions determining regioselectivity (Fisher et al., 1998). One
`might expect the models to work better with enzymes and substrates
`that show little preference in orientation, such as CYP1A2 (Higgins et
`al., 2001), CYP2B1 (Higgins et al., 2001), CYP2B4 (White et al.,
`1984), CYP2E1 (Higgins et al., 2001), and CYP3A4. It should be
`noted that, although these enzymes have been shown to be free of
`excessive orientation effects for some substrates, it is possible that
`other substrates will show strong orientation effects in any one of
`these enzymes. When binding/orientation effects become large, mod-
`
`els that account for binding effects on regioselectivity can be com-
`bined with the electronic models to give a more complete description.
`It is important to note that, although for certain enzymes the electronic
`model can stand alone to predict regioselectivity, it is doubtful that
`binding models alone, without an electronic model, can predict regi-
`oselectivity. Even strongly orienting P450 enzymes can give multiple
`products, and the electronic nature of each position must be known.
`Other considerations that need to be made when using this model
`are the general steric accessibility of the positions of interest. For
`example, octane is metabolized to 1-octanol, 2-octanol, and 3-octanol
`in the ratio of 1:23:7 (Jones et al., 1990). The model has been shown
`to predict the correct regioselectivity for the C-1 and C-2 positions;
`however, the electronics of the C-2 and C-3 position are essentially
`the same so the 23:7 ratio would not be well predicted. In this case, the
`C-3 position appears to be less accessible since it is buried in the
`center of the molecule. Similarly, C-4 hydroxylation is not seen with
`octane to any appreciable extent, whereas the electronic features are
`not significantly different from those at the C-2 position. Another
`feature of this type is ortho hydroxylation. For many compounds, the
`group adjacent to the ortho position will hinder metabolism at the
`ortho position. These types of steric shielding are independent of the
`P450 enzyme and could be added to the models presented here.
`Finally, it should be noted that no attempt has been made to
`externally validate these models. Although this would obviously be
`preferred, we felt that the datasets were too small to remove data for
`
`Apotex Ex. 1016
`
`

`

`Downloaded from
`
`dmd.aspetjournals.org
`
` at ASPET Journals on April 14, 2020
`
`12
`
`JONES ET AL.
`
`external validation. We are currently working on developing new
`datasets to validate these models. We did look at the internal consis-
`tency of the models by leave-one-out cross validation and found q2
`values of 0.79 for the combined model (data in Fig. 4). The cross-
`validated q2 for leaving out 20% of the data ranged from 0.76 to 0.81
`over 10 runs with different randomly chosen groups.
`In conclusion, computational models are presented that can predict
`the two major P450 pathways of metabolism, aliphatic and aromatic
`oxidation reactions. The models depend on experimental data to
`correct for the errors associated with the fast semiempirical methods
`used to predict each reaction mechanism. The computational models
`corrected by experimentally measured regioselectivities give the pre-
`dictive models described by eqs. 3 and 4. These models do a reason-
`able job at predicting the regioselectivity energetics and can be used
`in conjunction with models for the active site of the enzyme. Studies
`are underway in the application of these models to predicting human
`drug metabolism.
`
`References
`
`Atkins WM and Sligar SG (1987) Metabolic switching in cytochrome P-450cam: deuterium
`isotope effects on regiospecificity and the monooxygenase/oxidase ratio. J Am Chem Soc
`109:3754 –3760.
`de Groot MJ, Ackland MJ, Horne VA, Alex AA, and Jones BC (1999) A novel approach to
`predicting P450-mediated drug metabolism. CYP2D6 catalyzed N-dealkylation reactions and
`qualitative metabolite predictions using a combined protein and pharmacophore model for
`CYP2D6. J Med Chem 42:4062– 4070.
`de Groot MJ, Denkelder GMDO, Commandeur JNM, Vanlenthe JH, and Vermeulen NPE (1995)
`Metabolite predictions for parasubstituted anisoles based on Ab-Initio complete active space
`self-consistent-field calculations. Chem Res Toxicol 8:437– 443.
`Dewar MJS, Zoebisch EG, Healy EF, and Stewart JJP (1985) AM1: a new general purpose
`quantum mechanical molecular model. J Am Chem Soc 107:3902–3909.
`Fisher MB, Zheng YM, and Rettie AE (1998) Positional specificity of rabbit CYP4B1 for
`omega-hydroxylation 1 of short-medium chain fatty acids and hydrocarbons. Biochem Biophys
`Res Commun 248:352–355.
`Grogan J, DeVito SC, Pearlman RS, and Korzekwa KR (1992) Modeling cyanide release from
`nitriles: prediction of cytochrome P450-mediated acute nitrile toxicity. Chem Res Toxicol
`5:548 –552.
`Hanzlik RP, Hogberg K, and Judson CM (1984) Microsomal hydroxylation of specifically
`deuterated monosubstituted benzenes. Evidence for direct aromatic hydroxylation. Biochem-
`istry 23:3048 –3055.
`Hanzlik RP and Ling K-HJ (1990) Active site dynamics of toluene hydroxylation by cytochrome
`P450. J Org Chem 55: