Introduction

Top Abstract Introduction 3D-QSAR and Pharmacophores CYP Models Computational Models for CYP... Discussion References

By the end of the 1990s, several reviews had characterized the active site details and physicochemical properties of substrates for the major cytochrome P450 (CYP¹) enzymes. These reviews had been gathered from analysis of physicochemical data (Smith and Jones, 1992), as well as an analysis of the literature for three-dimensional quantitative structure activity relationships (3D-QSAR), protein homology modeling and pharmacophore modeling (de Groot and Vermeulen, 1997), nuclear magnetic resonance, and site-directed mutagenesis techniques (Smith et al., 1997a,b). There has also been some suggestion of substrate rules defined in the form of a decision tree that describes determinants of CYP specificity (Lewis et al., 1998, 1999a). More recent reviews have addressed how a small number of physicochemical descriptors can explain the discriminatory abilities of the human CYPs (Lewis, 2000a,b) and provide a wealth of information for the interested reader. These types of data are important because there is still no crystal structure for a human membrane-bound CYP, although a crystal structure of the rabbit CYP2C5 is now available (Cosme and Johnson, 2000; Williams et al., 2000). In the face of this, these techniques and approaches have led to our most recent discoveries and insights into this large family of enzymes. Understanding the substrate specificity of the CYPs is essential because of their clinically important role in the metabolism of xenobiotics and endobiotics (Wrighton and Stevens, 1992). With the development of computational approaches, i.e., 3D-QSAR techniques, we are faced with a more visual description of how molecules may bind CYPs as either substrates and/or inhibitors. These techniques are still in the very early stages of development; as such, we have yet to see truly universal acceptance of this methodology. In an attempt to rectify this situation, it is worth considering that the foundation of these tools, mathematical approaches to understanding metabolism, is not entirely a modern endeavor, as these approaches began in the late 1960s (Hansch et al., 1968; Hansch, 1972). As early as 1972, Hansch had already suggested that quantitative relationships could be generated for the lipophilic character of drugs and metabolic parameters. Later QSARs for oxidation by CYPs were correlated with hydrophobic and steric parameters of substrates (Hansch and Zhang, 1993; Gao and Hansch, 1996). The publications by this group were perhaps a sign that CYPs could be modeled mathematically as a first step before computational analysis. However, we have now shifted to modeling specific CYPs and using tools with increased functionality available to us. To some extent these types of classical QSAR studies have been continued and expanded upon by relating the findings to docking of molecules in homology models of CYPs (Lewis, 1997; Lewis and Lake, 1998). The focus of this review, however, does not include such qualitative mammalian CYP homology models discussed in detail by others (Szklarz and Halpert, 1997b, 1998; Dai et al., 2000), which can be used to study enzyme function (Szklarz et al., 2000) or enable suggested semiquantitative models (De Rienzo et al., 2000). This review intends to give an overview of the pharmacophore and 3D-QSAR models that have been used to describe P450s and indicate their varying degrees of success.

	3D-QSAR and Pharmacophores

Top Abstract Introduction 3D-QSAR and Pharmacophores CYP Models Computational Models for CYP... Discussion References

The development of computational tools has paralleled that of in vitro approaches to understanding and characterizing CYPs. One of the first visual 3D-QSAR computational approaches was comparative molecular field analysis (CoMFA) (Cramer et al., 1988), which enabled interpretation and understanding of enzyme active sites when a crystal structure was absent. However, it was not until in vitro drug-drug interaction studies were widely used (through the 1990s) in the pharmaceutical industry that a need arose for faster and cheaper approaches to determine this parameter. In this arena, computational approaches like CoMFA became useful in understanding requirements for CYP inhibitors (Fuhr et al., 1993; Strobl et al., 1993). Such computational 3D-QSAR techniques use relevant conformers of ligands to suggest functional groups, the geometry of structural features, and/or regions of electrostatic and steric interactions essential for activity or fit to the active site. The overall combination and 3D spatial distribution of physicochemical properties, the functional groups of the ligands, and a measure of binding site properties of an enzyme, such as the K_{m (apparent)} (Nelsestuen and Martinez, 1997), K_{i (apparent)}, or IC₅₀ itself, define a pharmacophore. The methodologies used for generating 3D-QSAR applicable to CYPs have been reviewed in detail elsewhere (Ekins et al., 2000b,c).

Until recently, few CYP binding or active site models had been generated using enzyme kinetic data, and these focused primarily on inhibition. Now, however, a considerable number of CYP pharmacophores have appeared in the literature, which presents us with the opportunity to review what is known about several CYPs based on such computational analyses.

	CYP Models

Top Abstract Introduction 3D-QSAR and Pharmacophores CYP Models Computational Models for CYP... Discussion References

CYP1A2. CYP1A2 is an inducible member of the CYP superfamily, which can be inhibited by some selective serotonin reuptake inhibitors (Brosen et al., 1993) and antiarrhythmics (Kobayashi et al., 1998). One of the first human CYP1A2 3D-QSAR studies was performed with SYBYL and ALCHEMY II software with 44 energy-minimized quinolone antibacterials (Fuhr et al., 1993), using the percentage of inhibition of caffeine 3-demethylation as the CYP1A2 specific activity probe (Fuhr et al., 1993). All 44 molecules were assumed by the authors to be competitive inhibitors, and the percent inhibition data were suggested to be similar to that which would be obtained if K_i or IC₅₀ values were calculated. This study showed that there was a lack of correlation between lipophilicity and inhibitory effect. Four pharmacophore features were identified and suggested to be involved in the binding of CYP1A2 inhibitors in the active site. These were two positive potentials and two negative potentials that were thought to be necessary for binding of the quinolone type CYP1A2 inhibitors in the active site (Fuhr et al., 1993). However, these authors did not state interfeature dimensions or try to predict the inhibitory potential of molecules excluded from the model. In contrast, a later theoretical study suggested that the methyl group present at the 8-position on the xanthine ring of some tri- and di-substituted xanthine inhibitors is important for CYP1A2 inhibition. This would also be within 3 Å of one of the three regions of negative electrostatic potential (Sanz et al., 1994). A QSAR analysis using INSIGHT II with data for the inhibition of caffeine N3-demethylation by 16 naturally occurring flavonoids showed that the volume to surface area ratio descriptor had the greatest effect. In addition, the sigma factor (Hammett coefficient of the B ring) and electron densities at the C4' and C3 atoms were also influential (Lee et al., 1998). Planar molecules with a small volume to surface area were the most potent inhibitors, while the number of hydroxyl groups was inversely proportional to inhibition, and glycosylation of the free hydroxyls decreased inhibition (Lee et al., 1998). A QSAR analysis of 12 heterocyclic amines using the COMBINE and GRID/GOLPE approaches enabled a comparison of structures of the ligand-protein complex with the structures of the ligands alone, respectively (Lozano et al., 2000). Both methods produced similar results and provided useful insights into positions of likely hydrogen bonding or placement of hydrophobic or bulky groups (Lozano et al., 2000).

CYP2A6. To date there has been no published CYP2A6 QSAR; however, the related mouse form, CYP2A5, has been studied. One group analyzed substrate requirements using a graphical method and concluded that bicyclic ring systems with an electron-rich moiety were essential for the 11 molecules analyzed (Tegtmeier and Legrum, 1994). A CoMFA study with 16 substrates and inhibitors of CYP2A5 generated from literature data suggested the importance of negative electrostatic potential close to a lactone moiety, steric effects around the methoxy group on methoxsalen, and positive electrostatic potential para to the methoxy group (Poso et al., 1995). When partial least-squares molecular surface weighted holistic invariant molecular (PLS MS-WHIM), an alignment-independent approach, was taken to modeling this same data set, size, positive molecular electrostatic potential, and hydrogen bonding acceptor were important in producing a statistically significant PLS model (Bravi and Wikel, 2000). As yet, there have been few studies reporting large amounts of K_i or IC₅₀ data for the related human CYP2A6 (Draper et al., 1997). Until a data set for CYP2A6 becomes available, we are unlikely to see further 3D-QSAR for this enzyme; however, these models will probably be similar to those derived for CYP2A5.

CYP2B6. Many examples of xenobiotics metabolized in part by CYP2B6 have been identified and described in more detail (Ekins and Wrighton, 1999). Recently a Catalyst hypothesis was generated from 16 K_{m (apparent)} values either generated by the authors or obtained from the literature. Four features were suggested to be necessary for substrate binding in the active site, namely, three hydrophobes and one hydrogen bond acceptor (Ekins et al., 1999c). The three hydrophobic regions present in the pharmacophore were located at distances of 5.3, 3.1, and 4.6 Å from the hydrogen bond acceptor, with intermediate angles of 72.8° and 67.6°. The Catalyst pharmacophore demonstrated a good correlation of observed versus estimated K_m values (r = 0.85). A preliminary version of this pharmacophore was independently suggested as being consistent with the homology model constructed by another group (Lewis et al., 1999b). In addition, the review of numerous QSAR studies on CYP2B proteins suggested the importance of hydrophobic and electronic properties in the binding of substrates (Lewis et al., 1999b), in further agreement with the aforementioned Catalyst pharmacophore. A second 3D-QSAR approach using molecular surface descriptors was also used (PLS MS-WHIM). This produced a model with a leave-one-out q² value of 0.67 and a more realistic five random groups repeated up to 100 times q² value of 0.61. Interestingly, the molecular surface properties, hydrophobicity, and hydrogen bond acceptor capacity were selected as important alongside unweighted and positive electrostatic potential (Ekins et al., 1999c). Both of these computational approaches were also used to predict a test set of five compounds. In four of five cases, the K_m value was predicted within 1 log residual. These results in themselves are important as Catalyst and PLS MS-WHIM represent two quite distinct modeling approaches. Within Catalyst, molecules are described in terms of pharmacophoric features, i.e., hydrogen bond donor or acceptor atoms, hydrophobic moieties, and positively or negatively charged groups. The final result is represented by the 3D spatial disposition of essential features for the activity. In contrast, MS-WHIM descriptors are alignment-independent and are aimed at capturing global 3D chemical information at the molecular surface level. The shape and the electronic properties of a molecule are, therefore, characterized in 3D space and condensed into a single numerical vector. A structure-activity relationship is derived by means of PLS, and the final model is presented in the form of a linear equation. The fact that both of these approaches selected similar features suggests some degree of concordance or intermethod validation, although both methods failed to predict a different molecule in the test set, which suggests a possible advantage of using multiple and different 3D-QSAR approaches with data derived for CYPs. To date there have been no reported pharmacophore models for inhibitors of CYP2B6.

CYP2C9. Smith and Jones (1992) originally proposed that CYP2C9 might be a target for pharmacophore modeling in 1992. This suggestion was followed by the first manually superimposed pharmacophore model of CYP2C9 in 1993, which speculated that the active-site contained a hydrogen bond donor or acceptor (Jones et al., 1993). A more detailed model based on the overlay of eight substrates and one inhibitor indicated that a hydrogen bond donor heteroatom was around 7 Å from the site of metabolism (Jones et al., 1996a). While this model was not statistically validated, it did indicate that overlays of substrates may provide a method to probe the active site of this enzyme.

CYP2C19. One group has focused on obtaining substrate structure activity relationships for the polymorphic CYP2C19 using inhibitors of omeprazole 5-hydroxylation (Lock et al., 1998a,b). Using mainly benzodiazepines which are N-dealkylated and 3-hydroxylated, it was suggested that these sites and the carbonyl group were important for inhibition. Electron-withdrawing groups were found to further decrease inhibition. As yet, the data for the 14 compounds used in these two studies have not been used to produce a published 3D-QSAR.

CYP2D6. Human CYP2D6 is a polymorphic member of the CYP superfamily and is absent in 5 to 9% of the Caucasian population as a result of a recessive inheritance of gene mutations (Mahgoub et al., 1977; Eichelbaum et al., 1979; Armstrong et al., 1994). This results in deficiencies in drug oxidations known as the debrisoquine/sparteine polymorphism, which affect the metabolism of numerous drugs. A diminished metabolism of these drugs is found in poor metabolizers, which have two nonfunctional CYP2D6 alleles, compared with extensive metabolizers with at least one functional allele. A relatively large number of small-molecule models has been derived for this particular human CYP isoenzyme, using a variety of substrates or inhibitors (Wolff et al., 1985; Meyer et al., 1986; Islam et al., 1991; Strobl, 1991; Koymans et al., 1992).

CYP2E1. CYP2E1 is involved in the metabolism of many toxic and carcinogenic molecules such as low molecular weight solvents and anesthetics. Early on, it was suggested that the active site was restricted due to the limited size of known substrates. A graphical model of the active site topology was derived from reactions of human CYP2E1 with phenyldiazene, 2-naphthyl, and p-biphenylhydrazine. This work indicated that the active site was open above the pyrrole rings A and D of the heme for a height of 10 Å (Mackman et al., 1996). A CoMFA study using rat in vivo data derived for 12 chlorinated volatile organic compounds essentially related to metabolism by CYP2E1 and CYP2B1 suggested the importance of gaining access to the active site relative to ligand-enzyme complementarity (Waller et al., 1996). This work was presented as an easily visualized schematic consisting of a long channel through which the substrate must progress before reaching the active site. The importance of electrostatic (long-range recognition), steric (substrate size), and hydropathy (substrate surface interaction) was demonstrated in the resulting QSAR. A recent study has described dissociation constants for ethanol, halothane, isoniazid, and pyrazole with a range from 0.011 to 167 mM; this study also modeled arachidonic acid in CYP2E1 (Smith et al., 2000), which seemed to agree with the previous hypotheses of a long hydrophobic access channel.

CYP3A4. Smith et al. have described in detail the CYP3A4 active site characteristics (as well as those of the other major mammalian CYPs) based on homology models built using soluble bacterial CYP structures as a template (Smith et al., 1997a). They also suggested that the binding of CYP3A4 substrates in the active site is due to lipophilic forces, based on evidence from octanol and cyclohexane partition coefficients (Smith et al., 1997b). Additionally, these authors have suggested that the flexibility of the conformation of the CYP3A4 active site indicated by many researchers may contribute to the overall diverse substrate selectivity of CYP3A4 (Smith et al., 1997b). Many other papers have focused on the active site of this CYP using homology models and site-directed mutagenesis. Structural requirements of CYP3A4 substrates have been suggested to include a hydrogen bond acceptor atom 5.5 to 7.8 Å from the site of metabolism and 3 Å from the oxygen molecule associated with the heme (Lewis et al., 1996).

CYP19 (Aromatase). The importance of CYPs that metabolize endogenous substrates can be demonstrated by aromatase, which catalyzes the metabolism of androstenedione to estrone, 16-hydroxyandrostenedione to estriol, and testosterone to estradiol via the aromatization of the A ring and the removal of the C19 methyl group (Oprea and Garcia, 1996). Some early QSAR on this enzyme used molecular mechanics and quantum chemical calculations to show that the C4 to C5 double bond on 46 derivatives of 4-androstene-3,17-dione and 5-androstane-3,17-dione enables the adoption of a readily hydroxylated conformation for these inhibitors of human placental aromatase (Bohl et al., 1988). Later work using 3D-QSAR with substituted dichlorophenyl aromatase inhibitors generated two possible pharmacophores resulting from molecular descriptors and multidimensional linear regression (Nagy et al., 1994). Both pharmacophores used the two aromatic rings to define a rigid, nonpolar molecular shape in space and also contained a hydrogen bond acceptor (Nagy et al., 1994). In 1996, two groups described CoMFA studies for steroidal (Oprea and Garcia, 1996) and nonsteroidal (Recanatini, 1996) inhibitors of aromatase. The former study analyzed 50 steroidal inhibitors using CoMFA and GOLPE to suggest two hydrophobic binding pockets in the C6 region of the steroid (Oprea and Garcia, 1996). One of these is large and located in the -face while another is smaller and located in the 6-position. The latter paper also used CoMFA for 29 molecules related to fadrozole and aligned on S-fadrozole. This model suggests two regions around the reference molecule and defines the importance of the -region; it also presents an idea of the size of the hydrophobic site (Recanatini, 1996). This study was later expanded to include tetralone derivatives to provide a training set of 49 inhibitor molecules and a further test set of eight molecules (Recanatini and Cavalli, 1998). It was found necessary to use two different aromatase inhibitors for alignment, suggesting that there may be multiple binding orientations for different classes of nonsteroidal aromatase inhibitors (Recanatini and Cavalli, 1998). Further diverse inhibitors of aromatase have been synthesized, such as the aryl-substituted pyrrolizine and indolizine derivatives (Sonnet et al., 2000), but it must also be considered that such molecules may be inhibitors for other CYPs. An example is the aromatase inhibitor anastrozole, which is a nanomolar inhibitor of human placental aromatase and a micromolar inhibitor of CYPs 1A2, 2C9, and 3A (Grimm and Dyroff, 1997). The degree of inhibition of these latter CYPs is, however, thought to be clinically insignificant.

CYP51 (14-Demethylase). Yoshida et al. (2000) recently reviewed the significance of CYP51, since this enzyme is widely distributed and conserved across biological kingdoms as the major sterol 14-demethylase. An initial understanding of the active site topology of this enzyme was obtained for Saccharomyces cerevisiae, which enabled a binding orientation for lanosterol to be proposed (Tuck et al., 1992). This first template type model proposed the importance of hydrogen bonding near the 14-methyl group. The first QSAR that appeared for fungal CYP51 used energy-minimized structures of 40 imidazole and triazole inhibitors of Candida albicans, aligned to derive a CoMFA model (Tafi et al., 1996). Two alignments were attempted, one overlapping azole rings only, and a second in which azole and phenyl rings attached to an asymmetric carbon atom were superimposed. The model derived from the second alignment showed that imidazoles were more active than triazoles and was predictive for 15 molecules excluded from the model, suggesting its utility in designing new inhibitors for this CYP (Tafi et al., 1996). A pharmacophore was recently derived for this enzyme in C. albicans using Apex-3D analysis and was followed by molecular volume analysis using INSIGHT. The best model consisted of 11 of 13 azole antifungal molecules (Talele and Kulkarni, 1999) and suggested that the N3 and N4 of the azole ring, the ethereal oxygen atom, and an aromatic ring centroid are essential in the binding site. This model was then updated with 18 of 20 molecules, and the alignment used with CoMFA to accurately predict the in vitro inhibitory activity values of 4 molecules was omitted from the model. In addition, the CoMFA model confirmed the importance of the original three-point pharmacophore suggested with Apex-3D (Talele et al., 1999). Small data sets have been used to generate classical regression-type QSAR models for CYP51 inhibition to indicate that there may be interactions between the azole ring and the heme of fungal CYP51. In turn, these studies were used alongside homology models of the fungal CYP (Lewis et al., 1999c). The utility of this type of QSAR for the commercial sector has also been reviewed for the production of azole-type agricultural fungicides (Fujita, 1997). A recent study used both CoMFA and Catalyst to describe the shape of the CYP51 binding pocket for targeting with herbicides, suggesting the importance of hydrophobic and hydrogen bond acceptor features (Bargar et al., 1999). With the recent publication of IC₅₀ data for the human CYP51 for ketoconazole and itraconazole (Lamb et al., 1999), it will be interesting to see the development of pharmacophores derived for mammalian data and how they vary from those described above for yeast, fungi, and plant CYP51. With the identification of this enzyme in tuberculosis (Bellamine et al., 1999), a potential therapeutic approach may be strengthened by using a pharmacophore in the design of inhibitors for this bacterial CYP. Applying computational approaches like those discussed as well as homology models [using mammalian CYP51 alignments (Ji et al., 2000)] may result in more selective therapeutic agents that do not inhibit the host CYP51. This potential for inhibition may result in interference with meiosis-activating sterol production and could possibly affect mammalian reproduction (Debeljak et al., 2000). Alternatively, inhibition of human CYP51 may be therapeutically useful (Harwood et al., 1999); therefore, modeling this may be valuable. Clearly, as more in vitro data are generated, pharmacophore models for this human CYP will not be far behind.

	Computational Models for CYP Inducers

Top Abstract Introduction 3D-QSAR and Pharmacophores CYP Models Computational Models for CYP... Discussion References

Not only is the prediction of CYP-mediated inhibitory drug-drug interactions important for the pharmaceutical industry, but CYP induction represents an additional drug-drug interaction that should be avoided. Computational modeling of CYP inducers is complicated because most of the induction literature relates to animal in vivo data, which in some instances can be modeled using QSAR analysis (Rekka and Kourounakis, 1996) or in vitro data (van der Burght et al., 1999). The appearance of data sets from human in vitro models such as those for CYP3A4 (Scheutz et al., 1996), nuclear hormone receptors pregnane X receptor (Kliewer et al., 1998; Lehmann et al., 1998; Moore et al., 2000), and CAR (Honkakoski et al., 1998) offers potentially larger data sets. This would suggest that we might be able to define physicochemical properties or molecular features necessary for induction [in the latter cases, binding to one or more nuclear hormone receptor site(s)]. However, the narrow range of fold induction with these data sets applicable to humans may be a limitation with the presently available modeling approaches such as those described in this review. The ability to computationally predict CYP3A4 induction, for instance, would be a great advance in throughput over the currently available in vitro methods using hepatocytes and functional and binding assays.

	Discussion

Top Abstract Introduction 3D-QSAR and Pharmacophores CYP Models Computational Models for CYP... Discussion References

The first review to discuss pharmacophores applied to CYPs reviewed the literature up to 1993 and suggested the likely difficulty in producing active site templates for other CYPs that might not possess specific ionic interactions (Korzekwa and Jones, 1993). However, the present review and previous discussions on predicting drug-drug interactions in silico (Ekins et al., 2000b) point to the increased interest and relative ease in modeling all the major human CYPs. In conclusion, we have shown that various 3D-QSAR models can be generated and that these models could be used to classify molecules for their likely ability to be CYP substrates or inhibitors (Table 1). The models reviewed here show that although there are common features present in all P450s modeled so far, the relative contributions vary dramatically between different P450 isoenzymes (e.g., relative contributions of electrostatic versus hydrophobic interactions and the importance of flexibility of the active site). In the future, CYP induction will probably be modeled to the same extent as we now attempt to predict drug-drug interactions for these enzymes. This pharmacophoric/3D-QSAR approach would naturally be a very useful tool for virtual drug discovery once the models were suitably refined to account for some of the poor predictions presently observed. Such an iterative approach to model building would provide CYP prototype models that could be widely applied in discovery alongside other in silico models for absorption, distribution, metabolism, and excretion/toxicology properties and bioactivity to predict molecules likely to avoid attrition during development.

Sean Ekins was born in Grimsby, England, and received the HND degree in applied biology from Nottingham Trent University (formerly Nottingham Polytechnic) in 1991. He received the M.Sc. degree in clinical pharmacology and the Ph.D. degree from the University of Aberdeen, Scotland. His doctoral research evaluated in vitro systems for metabolism, supervised by Professors Gabrielle M. Hawksworth and M. Danny Burke and sponsored by Servier Research and Development.

In 1996 he began work as a postdoctoral fellow at the Lilly Research Laboratories (Indianapolis, IN) in the laboratory of Dr. Steven A. Wrighton. His work initially involved in vitro characterization of recombinant and human CYP2B6 and its substrates. This project evolved into using commercially available computational three-dimensional-quantitative structure-activity relationship software to define molecular features important for CYP substrates and inhibitors. In 1998 he joined Pfizer, Inc. (Groton, CT) and worked with a team on in vitro and in silico approaches to predict drug-drug interactions. Since 1999 he has been a Senior Computational Chemist at Lilly Research Laboratories and continues to do research on computational approaches for predicting ADME (absorption, distribution, metabolism, and excretion) and toxicology endpoints along with academic and industrial collaborators.

Dr. Ekins is an Associate Editor for the Journal of Pharmacological and Toxicological Methods.

Marcel J. de Groot was born in Utrecht, the Netherlands. He received the M.Sc. degree in chemistry from Utrecht University and the Ph.D. degree in computational toxicology from the Vrije Universiteit Amsterdam. His doctoral research was supervised by Professor Nico P. E. Vermeulen, Dr. Gabriëlle M. Donné Op-den-Kelder, and Dr. Joop H. van Lenthe (Utrecht University) and involved developing computational models for biotransformation enzymes.

Beginning in 1997 he worked as an Assistant Professor in Computational Chemistry at the Vrije Universiteit Amsterdam, after which he joined the Computational Chemistry group (now MISD) at Pfizer Global Research and Development in Sandwich, UK. Within Pfizer, Dr. de Groot continues to focus on computational models for biotransformation enzymes and the prediction of ADME (absorption, distribution, metabolism, and excretion) and toxicological related issues.

Jeffrey Jones received the B.S. degree in medicinal chemistry from the University of Michigan and the Ph.D. degree from the University of Washington where he studied with Bill Trager.

He completed his post-doctoral research with Mo Cleland at the University of Wisconsin. He was a faculty member at University of Rochester prior to moving to Washington State University. His main research interests involve studies of P450 enzymes for drug design and benign synthesis.