Current Drug Metabolism - Volume 4, Issue 5, 2003

Details of the mechanics of drug processing such as the identification and characterization of enzymes and transporters involved in drug elimination and metabolism and the metabolites deriving therefrom along with their dispositional fate and biological activities provide important information for both clinical and predictive purposes. Ultimately, the molecular details of interactions between drugs and drug metabolizing enzymes and / or transporters along with a refined picture of their regulation by both endobiotics and xenobiotics inform our predictions about downstream occurrences subsequent to the exposure of drug-processing cells to xenobiotics. However, the quantitative aspects of drug processing also possess tremendous significance for explaining clinical outcomes and drug-drug interaction phenomena. Much effort is currently being directed toward the prediction of dispositional and pharmacokinetic (PK) features of new chemical entities (NCEs) such that NCEs with unfavorable PK parameters and / or a high propensity for drug-drug interactions (DDIs) might be identified early in the pre-clinical stages of drug development. This issue of Current Drug Metabolism has assimilated contributions from some of the leading scientists in the world who discuss contemporary thinking about quantitative predictions of drug metabolism / transport and DDIs. The models and paradigms discussed here are far-reaching, and encompass three dimensional molecular and pharmacophore modeling of both cytochrome P450 (CYP) enzymes and nuclear receptors based on sequence alignment with mammalian crystallographic templates; physiologically-based pharmacokinetic (PBPK) modeling including the description and application of a convective-dispersion model of the liver; entirely novel response surface modeling that can be applied to the design and study of the interactions of N drugs; and even the establishment of a human drug metabolism database (hDMdb) that takes account of substrate structures in three dimensions, quantitative characterization of in vivo biotransformation pathways in humans, and more precise correlation between those biotransformation pathways and pre-clinical drug metabolism data. Also considered in this issue are the predictions of complex events entailing the co-processing of substrates by both CYPs and transport proteins and the regulation of transporter and CYP activities by interacting agents. Though each of the articles reviews topic literature, there is also an effort to describe or synthesize recent data not published at the time that this issue went to press. Collectively, then, these contributions actually go beyond describing the contemporary state of affairs in predictive drug metabolism and DDIs, but provide an interesting glimpse into what will likely be viewed as contemporary in the months and years just ahead. A brief overview of what the reader will discover in this issue follows. Dr. David Lewis describes methods for calculating binding energies for substrate binding to human CYP enzymes based on first principles. Using a total of ninety substrates he shows that the overall binding energy between a substrate and CYP enzyme can be accounted for from desolvation or partitioning energies, hydrogen bonding energy, π-π stacking energy, loss of bond rotational energy, ionic energy, and loss of translational / rotational energy. Values for these energies can be computed subsequent to in-silico substrate docking with 3-D models for each CYP constructed from sequence alignment between each human CYP and the mammalian crystallographic template, CYP2C5. The values for binding energy computed from in silico substrate-enzyme docking correspond so closely with calculated binding energies derived from measured values of either KD or Km such that less than 3% of the variance between in silico estimated binding energies and experimentally-derived binding energies is unaccounted for in regression plots between them. For a more limited number of CYP2B6 substrates whose binding affinities are dictated principally by their lipophilicity, there is a striking correlation between the negative log Km and log P values, permitting the estimation of Km values directly from log P values. Such in silico estimates of Km may provide an early picture of whether substrates are likely to be high or low clearance substrates. An important issue to resolve is whether Km values derived from molecular modeling and first principles will faithfully represent Km values in instances when enzyme kinetics are not adequately described by simple Michealis-Menten kinetics. Dr. Timothy Tracy succinctly summarizes the conditions that can lead to mis-characterization of enzyme kinetics if the Michaelis- Menten model is applied to characterize atypical drug metabolism kinetics that are more accurately described by more complex models. Such model misspecification can lead to errant predictions of intrinsic clearance and ultimately to errors in scaled in vivo clearances. Simple Michaelis-Menten kinetics are shown to yield erroneous predictions of intrinsic clearance when enzyme kinetics are actually biphasic, when there is homotropic or heterotropic cooperativity, if substrate inhibition occurs, or if partial enzyme inhibition occurs. Not only will model-misspecification give a false picture of a drug's true clearance value, but it can also likely yield inappropriate inferences about DDIs. In the examples that Tracy presents herein for CYP2C9 the clearance estimates associated with such model misspecification are approximately twofold in magnitude. Failure to take account of atypical enzyme kinetics, now known to occur for most CYPs, probably accounts substantially for the poor clearance correlations between multiple probes for the same CYP enzyme; a problem that has dogged the drug metabolism community for more than a decade. Great strides have been made in recent years in the prediction of the clinical significance of DDIs from in vitro experiments. This is especially true for interactions characterized by the inhibition of the metabolism of one drug (victim) by another (perpetrator). Generally, the clinical likelihood or significance of such an interaction can be predicted from the Ki value of the perpetrator and the achievable in vivo concentration, [I], occurring with usual doses. A broad concensus about the interpretation of the clinical significance of DDIs for which these parameters can be estimated has been reached by scientists representing the European Federation of Pharmaceutical Sciences (EUFEPS, the US Food and Drug Administration (FDA), and the American Association of Pharmaceutical Sciences (AAPS).1 More recently a concensus opinion of a PhRMA subcommittee composed of members of the Drug Metabolism, Clinical Pharmacology, and Safety Assessment Technical Working Groups closely reflected the opinions of the international working group.2 For inhibitors of CYPs the PhRMA group uses Cmax for [I], and sets the likelihood of a clinically relevant inhibitory interaction as follows: 1) remote, if Cmax / Ki is less than 0.1; 2) likely, if Cmax / Ki is greater than 1.0; and 3) possible, for the intermediate values of Cmax / Ki. One of the most useful features of these three broad predictions of the clinical relevance of inhibitory DDIs is that those DDIs that will not be clinically relevant should effectively be predicted early in preclinical development. Moreover, since the published monographs about DDIs abound with cautions about putatitive DDIs based on qualitative observations in which a perpetrator drug is known to inhibit the metabolism of a victim drug in vitro, even those DDIs can now be revisited quantitatively. Those that fail the 0.1 criterion (Cmax / Ki) can be safely ignored and eventually culled and purged from DDI monographs and databases. This would be enormously helpful to clinical practitioners who depend on DDI monographs and databases. The devil is, as they say, in the details, however. Does Cmax denote total or unbound drug? Is it sufficient to use plasma or whole blood concentrations, or must tissue (e.g. liver) concentrations or inhibitor concentrations in the hepatic portal vein be used? These issues and others are addressed in two separate papers. In the paper by Chien et al., the poor quantitative prediction of metabolic inhibitors on drugs subject to sequential gut-to-liver first pass metabolism associated with the use of systemic concentrations of inhibitor at steady-state, [Iss], are discussed. Predictive outcomes vary wildly depending upon how the inhibitor concentration is incorporated into the predictive model, i.e. whether Cmax, Cavess, or a full PK concentration-time profile of the inhibitor is modeled. Another variable that is frequently overlooked is the intra-subject variability in clearance. If the purpose of a developmental DDI study is to develop accurate prescriptive advice, then Chien et al. recommend population PK DDI studies that encompasses data acquired from Phase 2 / 3 study populations, and include covariate effects on PK such as age, weight, and gender. However, accurate quantitative forecasting of DDIs must also take account of, in addition to the foregoing, organ perfusion rates of all clearing organs, bioavailability from gut and liver, protein binding, and tissue partitioning. Predictions about the magnitude of in vivo inhibition of victim clearance from in vitro experiments can be evaluated using unbound plasma concentration, total plasma concentration, or the product of total plasma concentration and liver / water or liver / plasma partition coefficient (Kp). Useful predictions about the magnitude of drug-induced change in the clearance of a victim drug can often, though not always be made if Ki values, both in vitro and in vivo, are referenced to unbound drug. This is especially true when in vitro Ki values are determined in microsomal systems.3 In these systems Ki values that are referenced to total drug concentration actually increase dramatically as the total protein content of the system increases. Even after referencing Ki values to unbound drug, however, in vitro Ki values may understimate in vivo inhibitory potency by several fold. The explanations for this disparity have been attributed to active drug transport, the presence of circulating inhibitory metabolites in vivo, and even mechanism-based inhibition. Reconciling in vitro-in vivo Ki disparities, and accounting for them a priori in in vitro DDI experiments remains a formidable challenge, and this is just one of several important issues that is comprehensively discussed by Venkatakrishnan et al. in this issue. Dr. Rene Levy and colleagues at the University of Washington have assimilated a substantial database of inhibitory druginteractions involving drug metabolizing enzymes and transporters (http: / / depts.washington.edu / didbase / ). They have drawn upon this database to evaluate the relationship between the dose of an inhibitor and the extent of enzyme inhibition in humans in vivo for both reversible and irreversible inhibitors, and found that most often the inhibition of both reversible and irreversible inhibitors is dose-dependent. Interestingly, having begun with a database of more than 4000 DDI pairs, in vivo Ki values (Kiiv) could only be computed for four (4) of them (0.1%)! And even for these DDI pairs the calculation of Kiiv is based on the underlying assumption that target enzymes are operating by Michaelis Menten rather than atypical kinetics. Thus a universal strategy for identifying, quantifying, and applying [I] / Ki values to predict DDI outcomes remains elusive, though if the questions raised by Chien et al. and Venkatakrishnan et al. are eventually experimentally addressed it may be possible to eventually reconcile in vitro and in vivo inhibitory DDI data for the purposes of making useful quantitative clinical predictions based on in vitro data. As difficult as it may be to make useful in vitro to in vivo (IVIV) predictions about the magnitude of AUC changes of a victim drug elicited by an inhibitory perpetrator, IVIV predictions of AUC changes by enzyme inducers is challenging in its own right. Mankowski and Ekins discuss the complex and, in some cases, overlapping roles of multiple orphan nuclear receptors in regulating the expression of drug metabolizing enzymes-in some cases multiple drug metabolizing enzymes-transporters, and even other nuclear receptors. In contrast to the paper by Benet et al., Mankowski and Ekins focus on the upregulation of orphan nuclear receptors, e.g. PXR and CAR, by receptor agonists. They describe the considerable progress that has been made in identifying the chemical features of ligands that will enable useful predictions about NCE binding to these receptors. However, they note that predictions of ligand binding are not necessarily identical to predictions of receptor agonism. The latter predictions are likely to depend on the pairing of pharmacophore modeling with results of reporter gene assays or other measures of receptor function. The same types of quantitative issues descend on IVIV predictions for drug trafficking mediated by transporters as for CYPmediated biotransformation. However, here the problems become even more thorny, since, unlike enzymatic biotransformation, transporter-mediated trafficking can be bi-directional since multiple transporters may be involved. Benet et al. evaluate the complexity of transporter and enzyme-coupled processing, and the effects of this co-processing on the extraction ratios for substrates both in the intestines and the liver in the presence of perpetrators that inhibit either P-glycoprotein function, CYP3A4 function, or both. Appropriate IVIV predictions that take account of transporter-enzyme co-processing require that both modeling and measurements include intracellular concentrations of the substrate, since these values will change as a function of alterations in transporter function. Moreover, the effects of Pgp and CYP3A on the intestinal extraction ratio are complimentary, whereas in the liver they are counteractive. Sorting out or predicting outcomes for DDIs requires that, for a victim drug, one knows a) whether it is a Pgp substrate; b) whether it is CYP3A substrate; c) whether it is a substrate for both CYP3A and Pgp. Superimposed on this issue is the xenobiotic-mediated regulation of CYP enzymes and transporters. A given xenobiotic (perpetrator) while inhibiting, for example, CYP3A4 may simultaneously inhibit Pgp. Or it may upregulate Pgp. In fact, a xenobiotic may be capable of upregulating and inhibiting a target CYP (or transporter). How might these phenomena confound our ability to predict some net DDI outcomes a priori? Many of the issues addressed in the individual papers appearing in this issue that have been referred to above are also tackled in the sweeping review of the field by Venkatakrishnan et al. Additional topics covered in their comprehensive overview of experimental and quantitative considerations associated with in vitro to in vivo (IVIV) predictions include scaling problems and methodological problems such as solvent use, non-specific binding, mechanism based inactivation, and probe selectivity. One picture that is beginning to emerge is that while it may be possible to identify and quantitate onerous PK characteristics and DDIs for some NCEs pre-clinically, thereby enabling “go” versus “no go” decisions about further drug development early in the process, there will undoubtedly be new drugs for which significant DDIs will, in fact, be missed (false negatives) in preclinical studies. Thus, failure to detect significant DDIs pre-clinically does not predict the absence of significant DDIs, even in the absence of any identifiable methodological problems. Even if a uniform set of industry / government standards could be eventually established for conducting both in vitro and in vivo DDI experiments, and for undertaking IVIV predictions, any such standards or guidelines would almost certainly be grounded in predicting the outcome of a single perpetrator drug on a single victim drug. But the reality is that an aging population is subject to increasing polypharmacy, and new strategies must emerge to enable us to consider DDI outcomes when there are multiple perpetrators and potentially multiple victim drugs. The sine qua non for evaluating the effects of agent mixtures in varying combinations has been the isobologram and related surface models. But these approaches, too, are limited to dealing with one perpetrator and one victim agent, i.e. two-agent or bivalent combinations. In this issue White and co-workers describe a new model for evaluating the interactions of N agents, i..e. multi-valent combinations such as those that are described in the review by Venkatakrishnan et al. The model, a nonlinear mixture response surface paradigm denoted as the nonlinear mixture amount model with zero amounts (NLMAZ), was actually developed to evaluate the optimum dose combinations for multidrug chemotherapy. In their paper White et al. carefully describe model development and its application to detecting synergism and antagonism over a range of dose combinations for three antimetabolite drugs; trimetrexate, raltitrexed, and LY309887 in HCT-8 human ileocecal adenocarcinoma cells. Outcomes over a range of doses for each of three agents are exhibited by color coding on a ternary plot in which dose combinations exhibiting synergism are depicted with symbol colors ranging from yellow to orange. Dose combinations that are antagonistic are represented by symbol colors ranging from orange to red. Thus, it is possible to discern for a three-drug combination potentially one set of dose combinations capable of eliciting synergistic effects, and another set of dose combinations yielding antagonistic effects. In the context of dispositional DDIs, then, this strategy would permit, for example, the identification of dose (or concentration) combinations that might, in fact, be used safely together, even though a separate set of dose (or concentration) combinations could not be used safely together. Finally, the paper by Paul Erhardt describes a Herculean effort to establish a human drug metabolism database (hDMdb) that has just recently begun. This effort is supported, in part, by the IUPAC, and when completed, will be available to the scientific community in much the same way as the human genome has been made available. Like existing drug-metabolism databases, the hDMdb will have a foundation of structure-metabolism relationships (SMRs), but they will likely be organized into tiers such as 2D structure, 3D structure, molecular similarity / dissimilarity indices, and fuzzy coordinate matrices. Superimposed on these SMR tiers will be a prioritization of metabolic pathways that takes account of a substance's distribution pattern and metabolism in different tissues. The challenges associated with creating multiple tiers of SMRs and prioritizing biotransformation pathways are discussed. It is interesting to note that one potential solution to these problems is to develop a sufficiently large dataset so that the predictions are strengthened simply by the large number of entries. The predictions of the hDMdb with regard to either expected metabolic outcomes or, for that matter, DDIs will therefore be inextricably linked to voluminous a posteriori information about in vivo human drug metabolism.

A straightforward methodology, based on first principles, for the estimation of human cytochrome P450-substrate binding energies is outlined, and the system has then been applied successfully to a relatively large dataset of P450 substrates totalling 90 compounds. The results of Quantitative Structure-Activity Relationship (QSAR) analysis on the same dataset of cytochrome P450 (CYP) substrates from the CYP1 , CYP2, and CYP3 families, involving a total of 90 compounds, agree favourably with the original analysis based on first principles, thus confirming the use of average values for hydrogen bond and π-π stacking energies, together with utilizing log P values as an estimation of desolvation energies. This method is based on a linear summation of the various contributary factors to the process, including: desolvation, hydrogen bonding, π-π stacking, restricted bond rotation and other energies relating to loss in translational and rotational energy. It is found that, for the majority of P450 substrates investigated, the first four terms are required for a relatively good estimation (R = 0.98) of the substrate binding affinity (ΔGbind) towards CYP1 and CYP2 enzymes. Consequently, it would appear that the loss in rotational and translational energy, which is thought to occur on substrate binding, apparently has little effect in most cases, possibly due to some degree of residual motion of the enzyme-substrate complex within the endoplasmic reticulum membrane. However, the appearance of a small constant term in the QSAR equation could possibly relate to an average loss in translational and rotational energy for the 90 compounds studied in this investigation.

The most common method for estimating a drug's in vivo clearance from in vitro data has involved using the classical Michaelis-Menten model to describe the observed in vitro data and from this estimate intrinsic clearance. This process assumes that the drug obeys standard enzyme kinetics that can be described by this model. However, the observation of atypical enzyme kinetics, particularly involving cytochrome P450 enzymes, has become relatively common, and occurrences have been reported with virtually all of the cytochrome P450 isoforms. Since predictions of a drug's in vivo clearance are made based on kinetic parameters observed during in vitro experiments (in vitro estimated intrinsic clearance), mis-identification of the drug's kinetic profile can lead to inaccurate predictions of intrinsic clearance. In addition to these potential ramifications of in vitro - in vivo predictions, information is becoming available concerning the in vivo implications of drugs that exhibit atypical kinetic profiles. This review will discuss the various types of atypical kinetic profiles that may be observed during in vitro experiments, hypotheses for why they may occur, examples of how a drug's intrinsic clearance may be mis-estimated when atypical kinetics are incorrectly assessed, and the potential ramifications of these issues for prediction of drug interactions. Furthermore, potential, artifactual causes of atypical kinetic phenomena will be described.

The prediction of metabolic drug-drug interactions should include quantitative attributes, such as variability in the study populations, and the results should be presented in terms of probability and uncertainty. The simple algebraic equations used to calculate one mean value for the extent of drug-drug interaction are adequate for qualitative or semi-quantitative risk assessment. However, truly quantitative predictions continue to fail. The success of drug-drug interaction predictions requires understanding of the relationship between drug disposition and quantifiable influential factors on the change in systemic exposure. The complex interplay of influential factors, including variability estimates, on successful prediction of drug interaction have not been systematically examined. Therefore, physiologically relevant models of metabolic drug-drug interaction will likely play increasingly important roles in improving quantitative predictions and in the assessment of the influential factors underlying the interactions. The physiologically-based approach, with stochastic considerations, offers a powerful alternative to the empirical calculation of mean values. In addition to quantitative estimation of the interaction for assessing probability of risk, a reasonably validated predictive model is useful for prospective optimization of study designs. As a consequence, the definitive clinical trial would yield more meaningful information to support dosing recommendations. This review focuses on illustrating the importance of an integrated approach to building useful models for prediction of metabolism-based drug-drug interactions in human subjects.

Growing interest in the prediction of in vivo pharmacokinetic data from purely in vitro data has grown into a process known as the in vitro-in vivo correlation (IVIC). IVIC can be used to determine the viability of new chemical entities in the early drug development phases, leading to a reduction of resource spending by many large pharmaceutical companies. Here, a convective-dispersion model was developed to predict the total hepatic clearance of six drugs using pharmacokinetic data obtained from in vitro metabolism studies in which the drug disappearance from suspensions of human cryopreserved hepatocytes was measured. Predicted in vivo hepatic clearances estimated by the convective-dispersion model were ultimately compared to the actual clearance values and to in vivo hepatic clearances that were scaled based on the well-stirred model. Finally, sensitivity studies were performed to determine the dependence of hepatic clearance on a number of physiological model parameters. Results reaffirmed that low clearance drugs exhibit rate-limited metabolism, and their hepatic clearances are thus independent of blood flow characteristics, whereas drugs with relatively higher clearance values show a more pronounced dependence on the flood flow properties of dispersion and convection. Absent a priori knowledge about the flow-dependent properties of a drug's clearance, the convective dispersion model applied to disappearance data acquired from cryopreserved human hepatocytes is likely to provide satisfactory estimates of hepatic drug clearance.

A comprehensive search of the literature was undertaken using the Metabolism and Transport Drug Interaction Database (http: / / depts.washington.edu / didbase / ) to evaluate the relationship between extent of inhibition and inhibitor dose. The search included reversible and irreversible inhibitors in studies conducted in the period 1966-2003. Only twelve inhibitors met the criterion of the search: study population exposed to more than one dose of inhibitor within a given study design. Six were reversible inhibitors: ciprofloxacin, enoxacin, felbamate, fluconazole, fluvoxamine and ketoconazole. The other six (cimetidine, diltiazem, disulfiram, paroxetine, verapamil and ritonavir) are considered irreversible inhibitors. Most of the AUC / Clearance data available for both types of inhibitors suggested evidence of dose-dependent inhibition. In the case of reversible inhibitors, the evidence of dose-dependent inhibition is consistent with a number of recent studies suggesting the determination of in vivo inhibition constants based on plasma concentration of inhibitor.

New chemical entities are routinely screened in vitro and in vivo for their ability to induce cytochrome P450s (CYP), other drug-metabolizing enzymes and possibly transporters in an attempt to more accurately predict clinical parameters such as drug-drug interactions and clearance in humans. Some of these potential therapeutic agents can cause induction of the metabolism of another molecule or auto-induction thereby increasing their own metabolism and elimination, as well as potentially any molecules metabolized by the same enzyme(s). Key CYPs in the 1A, 2B, 2C, and 3A families have all been shown to be inducible. It would be clearly advantageous to know the potential for a compound to induce drug metabolizing enzymes or transporters prior to clinical development, and many in vitro systems have been developed for this purpose. Newer computational technologies are also being applied in order to attempt to predict induction from the molecular structure alone before a molecule is even synthesized or tested. This review will cover the various in vitro and in silico methods developed for prediction of key inducers of CYPs and other proteins, as well as the limitations of such technologies and applications in the future.

As discussed in earlier articles, predictions of in vivo drug-drug interactions from in vitro studies is a subject of high interest with obvious therapeutic as well as economic benefits. Up until now little attention has been given to the potential interplay between metabolic enzymes and transporters that could confound the in vivo-in vitro relationships. Drug efflux by intestinal Pglycoprotein (P-gp) is known to decrease the bioavailability of many CYP3A4 substrates. We have demonstrated that the interplay between P-gp and CYP3A4 at the apical intestinal membrane can increase the opportunity for drug metabolism by determining bidirectional extraction ratios across CYP3A4 transfected Caco-2 cells for two dual P-gp / CYP3A4 substrates, K77 (an experimental cysteine protease inhibitor) and sirolimus, as well as two negative control, CYP3A4 only substrates, midazolam and felodipine. Studies were carried out under control conditions, with a P-gp inhibitor (GG918) and with a dual inhibitor (cyclosporine). Measurement of intracellular concentration changes is an important component in calculating the extraction ratios. We hypothesize that the inverse orientation of P-gp and CYP3A4 in the liver will result in an opposite interactive effect in that organ. In vivo rat intestinal perfusion studies with K77 and rat liver perfusion studies with tacrolimus under control conditions and with inhibitors of CYP3A4 (troleandomycin), P-gp (GG918) and both CYP3A4 / P-gp (cyclosporine) lend support to our hypotheses. These results serve as a template for predicting enzymetransporter (both absorptive and efflux) interactions in the intestine and the liver.

A flexible approach to response surface modeling for the study of the joint action of three active anticancer agents is used to model a complex pattern of synergism, additivity and antagonism in an in vitro cell growth assay. The method for determining a useful nonlinear response surface model depends upon a series of steps using appropriate scaling of drug concentrations and effects, raw data modeling, and hierarchical parameter modeling. The method is applied to a very large in vitro study of the combined effect of Trimetrexate (TMQ), LY309887 (LY), and Tomudex (TDX) on inhibition of cancer cell growth. The base model employed for modeling dose-response effect is the four parameter Hill equation [1]. In the hierarchical aspect of the final model, the base Hill model is treated as a function of the total amount of the three drug mixture and the Hill parameters, background B, dose for 50% effect D50, and slope m, are understood as functions of the three drug fractions. The parameters are modeled using the canonical mixture polynomials from the mixture experiment methodologies introduced by Scheffe [2]. We label the model generated a Nonlinear Mixture Amount model with control observations, or zero amounts, an “NLMAZ” model. This modeling paradigm provides for the first time an effective statistical approach to modeling complex patterns of local synergism, additivity, and antagonism in the same data set, the possibility of including additional experimental components beyond those in the mixture, and the capability of modeling three or more drugs.

Previous attempts to predict drug metabolism and drug-drug interaction possibilities by deploying databases that house established drug metabolism results have been only marginally successful. Consideration of some of the key issues and concerns derived from these efforts suggests that three major hurdles loom in front of using xenobiotic metabolism databases more effectively in the future. These hurdles include: the need for an improved treatment of chemical structure in three-dimensions (3D); a better quantitative accounting of competitive and complementary biotransformation pathways; and, the critical need for a comprehensive, human drug metabolism database (hDMdb) that can serve as a bio / chemoinformatic resource pertaining to drug metabolism and drug metabolism-related drug-drug interactions in general. Approaches that might be taken to traverse each of these hurdles are discussed herein where: the first involves maturation of chemical structures from simple 2D entries into more sophisticated 3D displays that can also account for interactions with relevant biological surfaces; the second involves a systematic, pair-wise comparison of various metabolic options in a statistically driven manner relative to chemical structure descriptors; and, the third involves mounting a hDMdb on the Internet in a user-friendly manner that it is available via a non-profit format.

In vitro models of drug metabolism are being increasingly applied in the drug discovery and development process as tools for predicting human pharmacokinetics and for the prediction of drug-drug interaction risks associated with new chemical entities. The use of in vitro predictive approaches offers several advantages including minimization of compound attrition during development, with associated cost and time savings, as well as minimization of human risk due to the rational design of clinical drug-drug interaction studies. This article reviews the principles underlying the various mathematical models used to scale in vitro drug metabolism data to predict in vivo clearance and the magnitude of drugdrug interactions resulting from reversible as well as mechanism-based metabolic inhibition. Examples illustrating the predictive utility of specific in vitro approaches are critically reviewed. Commonly encountered uncertainties and sources of bias and error in the in vitro determination of intrinsic clearance and metabolic inhibitory potency, including nonspecific microsomal binding, solvent effects on enzyme activities, and uncertainties in estimating enzyme-available inhibitor concentrations are reviewed. In addition, the impact and clinical relevance of complexities such as dosing routedependent effects, atypical multi-site kinetics of drug-metabolizing enzymes, non-cytochrome P450 determinants of metabolic clearance, and concurrent inhibition and induction, on the applicability and predictive accuracy of current in vitro models are discussed.