Simcyp Limited, Sheffield, UK 2. Eli Lilly, Windlesham, UK 3. Academic Unit of Clinical Pharmacology, University of Sheffield, UK

Transcription

1 Drug Metab. Pharmacokinet. 24 (1): (2009). Review A Framework for Assessing Inter-individual Variability in Pharmacokinetics Using Virtual Human Populations and Integrating General Knowledge of Physical Chemistry, Biology, Anatomy, Physiology and Genetics: A Tale of `Bottom-Up' vs `Top-Down' Recognition of Covariates Masoud JAMEI 1, Gemma L DICKINSON 2 and Amin ROSTAMI-HODJEGAN 1,3, * 1 Simcyp Limited, Sheffield, UK 2 Eli Lilly, Windlesham, UK 3 Academic Unit of Clinical Pharmacology, University of Sheffield, UK Full text of this paper is available at Summary: An increasing number of failures in clinical stages of drug development have been related to the effects of candidate drugs in a sub-group of patients rather than the `average' person. Expectation of extreme effects or lack of therapeutic effects in some subgroups following administration of similar doses requires a full understanding of the issue of variability and the importance of identifying covariates that determine the exposure to the drug candidates in each individual. In any drug development program the earlier these covariates are known the better. An important component of the drive to decrease this failure rate in drug development involves attempts to use physiologically-based pharmacokinetics `bottom-up' modeling and simulation to optimize molecular features with respect to the absorption, distribution, metabolism and elimination (ADME) processes. The key element of this approach is the separation of information on the system (i.e. human body) from that of the drug (e.g. physicochemical characteristics determining permeability through membranes, partitioning to tissues, binding to plasma proteins or affinities toward certain enzymes and transporter proteins) and the study design (e.g. dose, route and frequency of administration, concomitant drugs and food). In this review, the classical `top-down' approach in covariate recognition is compared with the `bottom-up' paradigm. The determinants and sources of inter-individual variability in different stages of drug absorption, distribution, metabolism and excretion are discussed in detail. Further, the commonly known tools for simulating ADME properties are introduced. Keywords: ADME; developmental pharmacology; drug development; drug discovery; drug-drug interactions; Monte Carlo simulations; mathematical modeling; pharmacogenetics; physiologically-based pharmacokinetics; pharmacokinetic; pharmacodynamic modeling Introduction Individual variability in dose-concentration relationshipsanditsimpactondoserequirementsisofrelevance to clinicians as well as scientists working in drug development. Regulatory organizations require information on covariates relevant to patient populations at the point of filing applications for new candidate compounds. The most relevant information, with implications for dose adjustment in certain sub-group of patients or avoiding prescribing the drug to some other sub-groups, is then incorporated in the labeling for marketed drugs. For many years, statistical analyses of data obtained from small parallel studies, together with population pharmacokinetic (POP-PK) assessment of larger Phase II and Phase III studies, have been the two foundations of initial covariate recognition efforts. These methods require the analysis of data from a diverse population of individuals. Classical pharmacokinetic studies, to assess differences between groups, involve intense sampling from individuals but employ few people in each sub-group of the population under investigation. Conversely, POP-PK studies require Received; December 26, 2008, Accepted; January 16, 2009 *To whom correspondence should be addressed: Professor Amin ROSTAMI-HODJEGAN, Royal Hallamshire Hospital, Room M129-M Floor, The Medical School, Beech Hill Road, Sheffield, S10 2RX, UK. a.rostami@sheffield.ac.uk 53

2 54 Masoud JAMEI, et al. substantial numbers of individuals who may have provided sparse samples only. Since the introduction of the POP-PK approach in the 1980s, knowledge on mechanistic pharmacology and covariate analysis has increased significantly. Understanding the mechanisms by which covariates, such as weight, age, sex, renal impairment, genetic makeup of drug metabolizing enzymes and concurrent medication, can influence pharmacokinetics provides the opportunity to pre-select covariates (or anticipate effects in groups of patients who could not be studied in an experimental clinical setting see later sections). Under the new paradigm of covariate recognition, using the first principles and mechanistic models of physiologically-based pharmacokinetics (PBPK), any clinical data even when theyareobtainedforthefirsttime become`confirmatory', rather than a step in the `learning' process. An additional point to consider is the fact that most data analysis approaches to studying covariate effects examine only the simple relationships where the effects on dependent pharmacokinetic (or pharmacodynamic (PD)) parameters are assumed to be increasing or decreasing monotonically as the independent parameter values change. As shown later, this may not be the case and some covariate effects might be much more complex. This review presents recent progress in predicting the fate of drugs in the human body and assessing inter-individual variations in pharmacokinetics using modeling and simulation. Such investigations can form the basis for designing more efficient clinical trials, with cost-saving implications. Moreover, these methods might provide the only means of estimating pharmacokinetic differences in certain sub-populations whose characteristics may not be captured during typical Phase II studies. These include, butarenotrestrictedto: (i) subjects with rare combinations of genetic variants of drug metabolizing enzymes and transporter proteins (whose recruitment into trials in adequate numbers is not straight forward); (ii) patients who receive varying combinations of drugs (where testing drug interactions for all possible permutations of co-administrations is not possible); or (iii) individuals who cannot be investigated solely for thepurposeofexploringthevariations(e.g. drugdrug interactions (DDI) in children when the nature of interactions may be different in pediatrics than adults, and nonetheless carrying out such drug-drug interaction studies in pediatric volunteers is fraught with ethical problems). Characterization of Pharmacokinetics during Drug Development Classical Procedure `Top-Down' Approach: The traditional approach to drug development consists of four discrete stages: (i) discovery; (ii) pre-clinical research and development; (iii) clinical research and development; and (iv) `post-marketing' pharmacovigilance. 1) Within clinical research and development, Phase I investigations are the first point where potential covariates of pharmacokinetics in humans can be investigated, despite the limitation that these studies only involve small groups of healthy (typically young male) volunteers. Phase II studies, which are carried out in patients, provide further information on covariates in addition to their primary aim of providing knowledge on the safety, tolerability, effectiveness and appropriate dosage of the drug in the target population. However, covariates are most commonly investigated in Phase III where the appropriate dosing schemes are evaluated and information to secure drug approval is gathered. It is now acknowledged that the speed and cost of drug development can be optimized if many of these seemingly separate stages run concurrently rather than sequentially. Nonetheless, procedures for each drug are still being built based on observed data (`top-down') rather than taking advantage of broader knowledge of the human body and the information gained from previous cases for drugs which share similar characteristics (`bottom-up'). Although the identification and quantification of covariates, particularly using POP- PK, is now seen as an integral part of drug development, determining covariates using this approach is not straight forward and complications caused by bias and competition between multiple variables are well known and described in the literature. 2) New Paradigm `Bottom-Up' Approach: Modeling and simulation of the processes that define the plasma concentration-time course of a drug namely, absorption, distribution, metabolism and excretion (ADME) is an indispensable tool in integrating available prior information and accelerating decision making. The key element of this approach is the separation of information on the system (i.e. human body) from that of the drug (e.g. physicochemical characteristics determining permeability through membranes, partitioning to tissues, binding to plasma proteins, or affinities towards certain enzymes and transporter proteins) and the study design (e.g. dose, route and frequency of administration, concomitant drugs and food). The success in this field has been growing in parallel to the availability of the in vitro systems which act as surrogates for in vivo reactions relevant to ADME. In vitro in vivo extrapolation (IVIVE) has become possible not only because of advances in the understanding of the extrapolation factors (physical chemistry, biology, physiology and genetics) but because of the ability to `integrate' such information using mechanistic models of the human body 3) combined with the power of computers. 4) These efforts also run in parallel to other scientific activities under the descriptive umbrella of `Systems Biology' as dis-

3 Virtual Human Populations in Assessing Covariates of ADME 55 cussed recently in a special issue of the journal Xenobiotica. 5) Although initiatives for `model based drug development' in the pharmaceutical industry (e.g. see Lalonde et al. 6) ) and other pharmaceutical organizations (e.g. EU- FEPS 7) ) have been strong, a more powerful catalyst for the `systems approach' has been the willingness, and often leadership, of drug regulatory bodies in adopting the methodology; this is highlighted in published reports such as the FDA's Critical Path Initiative. 8) Requirements to Bridge the `Bottom-Up' and `Top-Down' Approaches: As indicated above, there are three elements which define the outcome of an ADME study and covariates affecting the observations. These are the characteristics of the system (i.e. the attributes of the human body for each subject), the characteristics of the drug and the conditions of the study. For a simulation platform to be useful in identifying covariates, these three elements should be adequately separated but interact with each other during the conduct of a virtual clinical study (Fig. 1). Some of the difficulties and successes in building such a system have been described recently by our group. 4) At the top level are the populations datasets which are independent of any specific drug or trial design but contain all the elements for the compound and trial design datasets to interact with. These include, but are not restricted to: data on enzymes/transporters and their abundances, including genotypes, rates of synthesis and degradation [mainly in liver and intestine]; intestinal and stomach motility, intestinal surface area and fluid dynamics; circulating levels of plasma proteins and red blood cells; organ size composition and organ blood flow. Each of these parameters can be determined and stored as an information library for various target populations, each of them having a different composition in regards to sex, age, ethnicity, and genotypic makeup affecting enzymes or transporters. In addition, special populations can be defined for groups of interest such as obese, cirrhotic and renally-impaired patients. As indicated by Jamei et al., 4) the efforts required to gather information for each population is much greater than those needed for the generation of the basic models; the time and resource implications for creating such databases are often underestimated. However, once created, each population library can be used repeatedly for any drug under any study design. Moreover, libraries can be updated and expanded as the knowledge of the system improves. Lastly, multiple libraries can be used simultaneously creating more divergent populations by mixing patients from different groups. The second component of any virtual ADME study is drug-specific information such as affinity to plasma proteins, red blood cells or enzymes and the ability to inhibit or induce certain enzymes. These data are usually gathered during different stages of drug discovery and development. These compound-specific data can be logically divided into different categories such as those related to `physical chemistry', `oral absorption', `organ distribution', `organ elimination', `inhibitory' or `induction' potential. Obviously, some parameters may affect more than one ADME process. The intensity, detail and quality of the drug-specific data typically increase as a candidate moves from the selection screens in the discovery stage to the optimization screens in pre-clinical development. Inevitably, many of the data required at earlier stages may be calculated from in silico models rather than measured. 9,10) Readers are referred to a recent report by Emoto et al. 11) for a typical example of the use of such in silico alternatives. The study highlighted the issue of measuring non-specific microsomal partitioning and plasma protein binding; where none of the current technologies for measurements are compatible with high throughput screening and hence they are difficult to implement at early discovery. Finally, the study design element includes information such as the fluid or food taken with the given dose, dosing frequency and administration period (if multiple-dose study), duration of blood sampling, makeup of the population (e.g. sex, genotypes of certain enzymes, age, ethnicity) and concomitant use of other drugs. At any point, the ADME in a selected individual can be simulated within the timeframe defined by the trial design; the procedure is repeated until all randomly selected individuals in the study population have been investigated. Interaction between the above three elements provides a complex covariate matrix (Fig. 2) whenallthe small building blocks of the PBPK models are put together (see next sections). This `bottom-up' approach provides the flexibility to seamlessly change the study design (e.g. the number of individuals, or diversity in the composition of the population). Hence, the power of studies to recognize covariates can be investigated apriori with the aim of improved decision-making. Examples of incorporating prior knowledge of the system and in vitro values into the design of covariate assessment are shown for both the classical parallel group studies (effect of enzyme genetics on PK 12,13) ) and POP-PK (assessment of DDI 14) ). Compliance (/adherence) to dosage regimen is another aspect of the study that can be simulated. However, the covariates which determine the degree of adherence to a given dosage regimen in an individual patient are not clearly defined. 15) Nonetheless, simulating the consequences of non-compliance in virtual clinical studies, as a source of possible variability in exposure to drug and hence the pharmacological and toxicological effects, provides invaluable information on the design of dosage regimen (e.g. on so-called `forgiveness' of missing doses, or the strategies to overcome the non-compliance) )

4 56 Masoud JAMEI, et al. Building Blocks of ADME: Potentials for Propagation of Variability The exposure of an individual to a certain drug can be measured by the area under the concentration time curve (AUC). The AUC after administration through any non-parenteral route (such as an oral dose) is dependent on the proportion of the dose that is absorbed and is subsequently available in the systemic circulation. In the case of oral drug administration (the most common route for drug intake), this involves release of the drug from the formulation, passage through the gut wall and then through the liver. The bioavailability of the drug (F) together with the clearance (CL) and the dose of the drug (D) will determine the overall exposure (AUC) according to Equation 1: AUC= F D Equation 1 CL Total CL is defined as the volume of blood completely cleared of drug per unit time and encompasses clearance by the liver, the kidneys and biliary excretion (in the absence of re-absorption from the gut). Although exposure to the drug is determined only by the dose, CL and bioavailability, varying shapes of concentration-time profile can occur for a given exposure when the rate of entry (absorption rate, infusion rate etc.) and rate of elimination are changed. Elimination rate is a function of CL and distribution characteristics. A brief description of ADME processes that determine the overall pharmacokinetics is provided below. Determinants of Oral Drug Absorption: Bioavailability (F) is a term often used to describe the absorption of a drug. It is defined as the proportion of an oral dose of a drug which reaches the systemic circulation in intact form. It is dependent on a number of key factors which are described by the following Equation: F=fa FG FH Equation 2 fa is the fraction of the dose which enters the gut wall (the remaining drug may be lost by decomposition in the gutlumen;itmaynotbereleasedfromtheformulation and remain in solid form; or it may become soluble in gut lumen but fail to permeate to the gut wall). FG is the fraction of drug which escapes metabolism in the gut wall and enters the portal vein, and FH is the fraction of the drug that enters the liver and escapes metabolism, thus entering systemic circulation. Overall bioavailability can be assessed by comparing the AUC values following oral and iv administration after correcting for any dose differences. Hepatic first pass effect can be estimated after decomposing the systemic clearance (iv administration) to its hepatic and renal components. However, estimating FG and fa from ordinary clinical data is not possible and many reports in the literature erroneously refer to the composite function of `FG fa'asifitrepresentsonly fa. Obviously, the latter could be true only if there is no gut wall metabolism at all. Current models describing the FG might be less mechanistic than other models however they have been useful in linking some of the in vitro data to clinical observations. Readers are referred to a recent report where the FG was described using an operational model. 19) In this so-called `QGut model' the FG was described using a flow term (QGut) which is a hybrid of both permeability through the enterocytes membrane and removal from the serosal side by villous blood flow as shown in the following equation: QGut FG= QGut+fuG S n ; CLuint, G «j j=1 wher QGut= Equation 3 Qvilli+CLperm$ where CLuint, G is the intrinsic metabolic clearance in the gut by the J th route based on unbound drug concentration, fug is the fraction unbound in the gut, CLperm is a clearance term defining permeability through the enterocytes, and Qvilli is actual villous blood flow. Determinants of Drug Distribution throughout the Body: Distribution refers to the reversible transfer of drug from one location to another within the body. Distribution of drugs to and from the blood and other tissues occurs at various rates and extents. Several factors areresponsibleforthedistributionpatternofadrugwithin the body over time. Many of these depend on the nature of the drug such as its ability to cross membranes, bind to plasma proteins, partition into red blood cells, tissues or fat, and its specific affinity to influx or efflux transporter proteins. Other factors determining the distribution behavior relate to characteristics of the individual such as the perfusion rate of different tissues by blood, the concentration of plasma proteins, hematocrit, body composition, tissue density, and genetic variants of transporter proteins. If any of the above elements show time- or dose-dependence then the dosage regimen may affect the pattern of the distribution observed at a given time after drug administration. Typical examples of timedependence involve the induction of transporter protein expression and non-retrievable inhibition (mechanismbased inactivation), whilst dose-dependent non-linearity is often due to saturation. Volume of distribution is the manifestation of drug distribution within the body and it influences the elimination rate and maximum exposure (Cmax). Volume of distribution together with clearance determines the rate of decline in plasma drug concentrations (elimination rate) the higher the volume, the longer the residence time in the body and vice versa. Since the proportion of the drug in different tissues changes with time (and the tissue-drug concentrations are not necessarily moving in parallel),

5 Virtual Human Populations in Assessing Covariates of ADME 57 Fig. 1. Schematic showing the principal elements of a population-based simulation platform Information on the system (i.e. the human body) is separate and can be used repeatedly for various drugs under different study designs (i.e. trial data). Although the arrows from the system to the central building block of the model (in vitro in vivo extrapolation [IVIVE] and physiologically-based pharmacokinetics [PBPK]) suggest a one-way flow, it should be noted that drugs may well influence the initial (baseline) parameters of the system. This happens through mechanisms such as induction of proteins or enzyme synthesis, and/or interference with degradation rate (acceleration or stabilization) of proteins/enzymes. Therefore, the level of enzymes and receptors should ideally be defined as a dynamic balance between the synthesis and degradation (rather than fixed static values) so any feedback effect following drug treatmentcanbeaccommodatedinrealtime(forfurtherdetailsseejameiet al ) ). Fig. 2. Overview of the relationships between covariates affecting ADME When building virtual human populations for ADME simulation, the composition of the study group is initially considered with respect to age, sex and ethnicity, plus genetic makeup of enzymes and transporter proteins in the target population. However, each of these factors influences multiple elements of ADME creating highly non-linear and non-monotonic relationships. The sensitivity of each pharmacokinetic parameter to a potential covariate depends on the nature of drug and the balance of sensitivities to elements within the network. As various drugs differ in their sensitivity to these elements, covariates of pharmacokinetics vary and a `one size fits all' solution cannot be assumed Simplistic assumptions for covariate analysis based on purely statistical models are a major shortfall for the current `top-down' data analysis in finding covariates. Prior assessment of covariates ensures that the most relevant factors and the most suitable covariate models are considered during clinical studies.

6 58 Masoud JAMEI, et al. volume of distribution is not a fixed term and it also changes with time. The V ss (volume of distribution at steady state) is considered when the ratio of drug in various tissues has reached equilibrium. It is seen as a `purer distributional term' 20) since other volume terms (such as central, Vc, andterminal,vz, distribution volume) can be affected by the relative speed of drug elimination and distribution. Traditionally, V ss has been calculated using the following formula after a drug has been administered intravenously (and assuming AUC is based ON drug concentrations in plasma): Vss= D MRT Equation 4 AUC where D=dose and MRT=mean residence time. However, this is an over-simplistic view of the processes involved. 21) Physiologically, volume of distribution is determined based on individual characteristics which go beyond simple links to body size, 22) as described by the following equation: V ss =Vp+Ve E:P+S t Vt Kp, t Equation 5 where Vp, Ve and Vt are volume of plasma, erythrocyte and tissue, respectively, and E:P and Kp, t are the relative drug concentrations in erythrocyte and tissue to plasma. It is clear from this equation that the characteristics of individuals (composition and size of tissues) can be separated from those of the drug (affinity to red cells or certain components of tissues). Determinants of Drug Metabolism: The majority of drugs currently on the market are lipophilic and metabolism is a major route of elimination from the body. 23) Understanding and fully characterizing metabolic routes helps with the early identification of potential covariates such as genetic polymorphisms of dug metabolizing enzymes, effects of environmental inducers or inhibitors, and any metabolic DDIs. However, it should be noted that overall metabolic clearance is not usually a simple linear function of the organ capacity (i.e. intrinsic clearance) but it is also dependent on the delivery of the free drug to the site of metabolism. Thus, hepatic clearance is determined by hepatic blood flow, plasma protein and red blood cell binding, and the effects of influx into or efflux from hepatocytes. In vivo intrinsic organ clearance can be extrapolated from a variety of in vitro systems using scaling factors as described by Barter et al. 24) and according to the procedure described by Rostami-Hodjegan and Tucker 3) : (a) using recombinantly expressed systems: CLu H,int = «n n S j=1 Ø S i=1 ISEF ji Vmaxi(rhEnzi) Enzjabundance Km i(rhenz j)»$ MPPGL Liver weight Equation 6 Where there are i metabolic pathways for each of j enzymes; `rh' indicates recombinantly expressed enzyme; Vmax is the maximum rate of metabolism by an individual enzyme; Km is the Michaelis constant; MPPGL is the amount of microsomal protein per gram of liver, and ISEF is a scaling factor that compensates for any difference in the activity per unit of enzyme between recombinant systems and hepatic enzymes. 25) (b) using human liver microsomes: CLuH, int=glu int(per mg Microsomes) MPPGL Liver weight Equation 7 (c) using human hepatocytes: GLu H,int =GLu int(per millions Hepatocytes) HPGL Liver weight Equation 8 where HPGL is hepatocellularity (millions of hepatocytes per gram of liver). These indicate the possibility of accommodating the covariate effects on metabolism which stem from ethnic/genetic differences (variability in the expression and activity of individual enzymes see Inoue et al. 26) ), specific diseases (e.g. liver cirrhosis and renal impairment see Nolin et al. 27) ), environmental factors (e.g. induction by smoking 28) )andage. 29,30) Some of these effects may influence more than one parameter at a time, for example, the ontogeny of enzymes occurs in the form of an agerelated change in the expression of enzyme per milligram of microsomal protein 31) ; however age also influences both MPPGL 29) and liver size. 30) This makes the `age-cl' relationship a drug-specific and complex matter with no `one size fits all' simple model (interested readers are referredtojohnsonet al. 31) for further information and examples). To determine whole organ clearance, estimated intrinsic clearance is combined with other determinants of CL using a variety of models. 32) The basic models (`wellstirred', `parallel tube', `dispersion') differ with regard to assumptions about the concentration gradient of drug within the liver, but in their simplest forms they all assume that the passage of the drug from the blood into the liver is perfusion-rate limited and that only unbound drug crosses the cell membrane and is available to be metabolized. Thus, they identify hepatic blood flow (QH), the fraction of drug unbound in blood (fub) and intrinsic metabolic clearance (CLuH, int) as the primary determinants of net hepatic blood clearance (CLB, H). On this

7 Virtual Human Populations in Assessing Covariates of ADME 59 basis, the `well-stirred' model, for example, predicts hepatic blood clearance and availability (FH) fromequations 9 and 10, respectively 33) (note: hepatic plasma clearance can be obtained by multiplying blood to plasma drug concentration ratio to terms in right hand side of E- quation 9; fub canbeexpressedintermsoffudividedby the blood to plasma concentration ratio of the drug): CLB, H= QH fub CLuH, int QH+fuB CLuH, int QH Equation 9 FH= Equation 10 QH+fuB CLuH, int An understanding of the variability in each of the primary determinants is necessary to predict the overall variability in hepatic drug clearance. Co-variation of blood flow with body surface area as well as age is well known, 34) and the effects of additional environmental (design) influences, such as eating, 35,36) posture and physical activity 37,40) can be estimated in these models prior to conducting any clinical studies. Similarly, hemodynamic changes in disease state (e.g. cirrhosis) may be incorporated into the prediction of variability, although the overall impact often depends on the multitude of other changes occurring in plasma proteins and red cells (Johnson et al., in preparation) as well as intrinsic clearance. Recent publications have examined the incorporation of the dynamics of binding and uptake (influx) information 41 43) and attempted to incorporate the knowledge of transporters and their interplay with drug metabolizing enzymes. Determinants of Drug Excretion: All drugs are ultimately removed from the body, either as metabolites or in their unchanged form. The primary route of secretion is through the kidneys and urine, although excretion may also occur via the biliary route and be considered as a true elimination when there is no re-absorption occurring in the intestine. Compromised renal function may affect the pharmacokinetics of a drug if urinary excretion is a substantial contributor to overall elimination. Drug characteristics which determine the extent of renal elimination include physical chemistry (lipophilicity and ionization), 44) plasma protein and erythrocyte binding 45) and affinity to certain transporter proteins in the kidney. 46,48) These mainly affect the fractional tubular re-absorption (FRe-abs), glumerular filtration or active secretion (CLuSec) of the drug which are summarized in Equation 11 [after Levy 45) and Jank ¹u 49) ]: CL R =Q R «fu B GFR Ø 1- fub GFR QR QR» Ø QR fub CLuSec, int (1-FRe-abs) Equation 11 QR+fuB CLuSec, int»$ Some of the `system related' parameters are self-evident from the Equation 11, such as glumerular filtration rate (GFR; as assessed by markers such as creatinine clearance) and renal blood flow (QR). However, there are other aspects of the system such as urine ph and urinary flow that influence the fractional tubular re-absorption (FRe-abs). The equilibrium between the drug residing in the erythrocytes and unbound concentration in plasma may not be rapid. Moreover, a proportion of erythrocytes might be separated off by `plasma skimming' and shunted into the renal veins without contacting the renal tubules 50) which may justify replacing QR and fub with renal plasma flow and fraction unbound of drug in plasma, respectively. 45) Any renal metabolism might be incorporated in equations such as Equation 11 by adding similar functions to that of the active secretion term. However, the sequential or parallel nature of such metabolic elimination relative to the active secretion is difficult to determine experimentally. An added complexity in assessing covariate effects for renal impairment is the correlation between kidney impairment and the expression of metabolic enzymes and transporters in the liver. 27) Determinants of Change in ADME with Co-administration of Different Drugs and Food: Drugdrug interaction has been identified as significant covariates of observed pharmacological response for a long time. 51) However, mechanistic understanding of the changes to drug effects following co-administration of drugs, and separation of the type of interactions based on the pharmacokinetic and pharmacodynamic related nature of them has taken place only in recent decades. 51) This mechanistic view has paved the way to using in vitro systems to understand and anticipate drug-drug interactions prior to conducting clinical studies. Attention to co-administration of drugs as a source of variability has grown after a number of high-profile problems with metabolic drug-drug interactions (mddi). Traditional approaches only involved the assessment of drug combinations which were more likely to be co-administered, or combinations which involved commonly used drugs with narrow therapeutic index. More recently, a POP-PK approach has been used to determine the effects of comedication 52) although the power of such studies is questioned when negative associations are reported. In other words, lack of any statistical significance for a given combination could be interpreted either as an indication of `no interaction' or it may indicate the inability of the study to identify a difference which stems from a failure in the power of the study. 14) Recent publications by the regulatory authorities 53,54) guide researchers towards a more mechanistic approach and rationalization of any investigations into the effects of co-administered drugs. This trend has been facilitated by the wider availability of programs and databases to assess the likelihood of interactions using relevant in vitro data 55,59) and help design

8 60 Masoud JAMEI, et al. the most appropriate study to identify worse case scenarios and satisfy regulatory requirements in understanding the theoretically conceivable effects. 60) Examples of incorporating variability into predictions of mddi using first principles are given in our earlier report 55) and point, among other factors, to genetic variability in non-inhibited pathways. This has since been reviewed by Collin and Levy 61) in relation to approved drugs on the market. Assessing any changes in drug exposure following co-administration of a competitive inhibitor of a metabolic pathway relies on understanding the proportional metabolism of the `victim' substrate via the inhibited route (fm) and the potency of the circulating concentrations of the inhibitor ('perpetrator') as described by the following equations: AUC (inhibited) AUC (uninhibited) = 1 n S (fmj Fold Change in CLuint, j=1 j)+ Ø 1-S n j=1 fmj» Equation 12 Fold Change in CLu int, j = 1 Equation 13 1+S p [I k] k=1 Ki k where fmj is the fraction of substrate clearance mediated by the inhibited metabolic pathway `j' and CLuint j is the intrinsic metabolic clearance of the substrate down pathway j. The fold reduction is defined here under multiple (`p') competitive inhibitors acting via the same mechanism to inhibit enzyme `j' where [Ik] is the concentration of inhibitor `k' at the enzyme site, and Ki k is the inhibition constant for inhibitor `k' obtained from in vitro studies after accounting for non-specific binding. If we consider that variation in fm depends on other metabolic routes, and that many of these routes are influenced by genetic polymorphism, then it is intuitive that, even at a fixed level of a given inhibitor ([I]), large variability in the extent of any significant mddi is likely to be observed between different individuals. Conversely, the circulating concentration of the inhibitor (and hence concentration at the active site) is subject to variability in all ADME processes which govern pharmacokinetics. Many of these variations can be predicted. In the case of induction, the fold change in activity of the induced enzyme depends on circulating concentrations of the inducer [I] (which could be variable in each individual); however the overall impact on AUC is also dependent on the fractional metabolism by the pathway (fm): AUC (induced) AUC (un-induced) = 1 n S (fmj Fold Change in CLuint, j=1 j)+ Ø 1-S n j=1 fmj» Equation 14 Fold Change in CLu int, j =1+ Ø Indmax [I] Equation 15 IndC50+[I]» where [I] is the concentration of the inducer at the site of effect, Indmax is the maximal increase in the level of induced enzyme (measured as a fold of the un-induced value) in the presence of a high concentration of inducer and IndC50 is the concentration of inducer associated with half maximal induction. More interesting types of interaction, that influence the system rather than directly affecting the interplay between the `victim' drug and the enzyme, are mechanismbased (suicidal) inactivation (MBI) of enzymes 62) and induction. 63) In the case of MBI, the fold reduction in clearance can be defined as the rate of inhibitor-related degradation relative to the natural degradation rate (kdeg)ofthe enzyme. The latter is an intrinsic parameter of the system (human body), rather than a drug dependent parameter. 64) The following equation represents a simplistic, pragmatic calculation for assessing the steady-state impact of MBI, although a more mechanistic prediction requires differential equations incorporating degradation rate which determines the dynamics of enzyme synthesis and elimination. 65) kdeg Fold Change in CLuint, j= Equation 16 k deg + [I] kinact [I]+K I where kinact is the maximum degradation rate constant in the presence of a high concentration of inhibitor and KI is the concentration of inhibitor associated with half maximal inactivation. One should note that all the above equations are only relevant for drugs undergoing linear `firstpass' and `systemic' hepatic metabolism according to the `well-stirred' model of hepatic elimination as they do not account for any inhibition of `first-pass' metabolism in the gut wall, transient plasma binding displacement during the absorption phase and its effect on hepatic `firstpass' metabolism or any variation in inhibitor concentration with time. Nonetheless, these, as examples, demonstrate the opportunities where inter-individual variability can be considered in predictions and the likely outcome of clinical studies. Further, complexities related to gut wall interactions can be added to these predictions 55) and some simpler, more pragmatic approaches involving full inhibition of gut first pass have also been described. 66) Concomitant food intake can sometimes directly affect adrug(e.g. chelating); however, the effect that food has

9 Virtual Human Populations in Assessing Covariates of ADME 61 on the system (human body), can have a significant impact on ADME depending on the characteristics of the drug. For example, changes in stomach ph and gastric emptying rate occur after food intake. These may or may not affect the bioavailability of drugs depending on the dosage form (solid or solution) and physicochemical characteristics (ionization status, lipophilicity and permeability). Similarly, changes in hemodynamics of intestinal blood flow or gut wall enzymes may affect the gut first pass of some drugs. Variability in the System Components (Human Body): Relevance to ADME The following section describes the available information on some of the system components which are required for the `bottom-up' approach when attempting to identify covariates. These are organized according to the building blocks of pharmacokinetics which were described above. In general, the information can be divided into `known-knowns' [sets of data which are known to be relevant for identifying covariates of pharmacokinetics which are available to integrate into models] and `knownunknowns' [sets of data which are known to be relevant but our current knowledge of them is `sparse' or`nonexistent']. It should be noted that there might be other relevant variables, of which we currently do not appreciatetherelevance[i.e. `unknown-unknowns']. As Benet and colleagues indicated over a quarter of century ago, ``at any point in the history of health care, our knowledge was considered to be quite extensive; however, in perspective, the knowledge of yesterday seems to have been very limited, just as today's knowledge can be expected to seem one day as such''. 67) Thus, it is of fundamental importance to note (a) the `evolving' nature of the knowledge that applies to the concepts of physical chemistry, anatomy, physiology, biology, genetics (proteomics) and epidemiology at any given time (our available databank) and (b) the crucial role of the framework (i.e. mathematical models) which integrates the information. Sources of Variability in Absorption: As described previously, absorption is affected by three parameters, fa, FG and FH, all of which are sensitive to inter-individual differences. Formulation-related aspects of drug release can be studied in vitro. These may include dissolution and solubility in aqueous solutions of different ph, simulated gastric fluid, fasted simulated intestinal fluid, fed simulated intestinal fluid; particle size measurements; and disintegration time. However, these physicochemical characteristics interact with the system-related factors. Hence, variability in fa is determined both by the formulation and physicochemical attributes of the drug and by variations in gastrointestinal (GI) motility and ph. The reported large variation in motility and residence time in the intestine (e.g. by Yu and co-workers 68) )(Fig. 3) mayormay not propagate into pharmacokinetic profiles 69) depending on drug and formulation characteristics. Thus, a poorly soluble and/or a low permeability drug, or a sustained release formulation may lead to more variation in bioavailability. The residence time of a drug in the stomach is an important factor determining the initiation of oral drug absorption.insomecasesitcanevendeterminetherate and extent of absorption. Stomach residence time can be influenced by elements of study design such as the volume of fluid administered with the solid dosage forms, concomitant food intake, 70,71) and composition of drug Fig. 3. Drug dependent propagation of physiological variability in the attributes of GI tract to drug bioavailability Part (A) illustrates the reported wide variation in motility of the GI tract (Yu et al ) ). This may or may not propagate to the estimated fractional absorption after oral administration of a drug solution depending on its permeability as shown in part (B) (taken from Jamei et al ) ). P eff indicates human gut wall permeability of the drug and T si refers to the intestinal transit time shown in Part (A). Similarly, formulation attributes (slow release) or dissolution characteristics may also determine whether the variability in transit time will affect the variability of bioavailability or not. For example, large variation in small intestinal transit time may have very little effect on the variability in bioavailability of an immediate release formulation of a soluble compound that is permeable through the gut wall. Conversely, clinical study conditions which are associated with more variable transit times would be of concern when designing studies for sustained release, sparingly soluble, low permeable drugs.

10 62 Masoud JAMEI, et al. formulation. 72) Our knowledge of various factors affecting gastric emptying is fairly extensive 73) and has been further expanded by advances in the area of imaging. 74) Nonetheless, integration of such information for apriori identification of the potential covariates has been less than optimal and many pharmaceutical companies go through unnecessary cycles of clinical studies involving formulation optimization without attention to the feasibility of reducing inter-individual variability and the source of such variation. It is known that stomach emptying could be a unique characteristic of each individual. This has been assessed by using repeated bioavailability studies on the same formulation under similar conditions (e.g. see Mojaverian et al. 75) ). However, in contrast to metabolic capacity or renal function, which can be assessed using biomarkers such as genotype or creatinine clearance, respectively, there are no biomarkers to determine the rate of stomach emptying in each individual prior to conducting the bioavailability study. The ph variation in the stomach and small intestine can influence the dissolution of solid dosage forms and affect the release of drugs from enteric coated formulations. In addition, ph may influence gut wall permeability by altering the balance between ionized and nonionized moieties. There is well-documented evidence for inter-individual 76) and inter-occasional 77) variation of ph throughout the GI tract as shown in Figure 4. Itisalso known that food and its composition, 70,78,80) pathophysiological conditions, 81) and disease (e.g. AIDS) 82,83) can alter the ph in the GI tract. The large inter-individual variation in residence time in the small intestine can propagate to oral bioavailability for certain groups of drugs (e.g. sparingly soluble, poor permeability, sustained release formulation). Thus, all conditions influencing intestinal transit times 71) may act as a determinant of bioavailability for such drugs. Study design (e.g. staggering the dose intake and food intake) is reported to be influential for pre-feed dosing where small-intestinal transit times were significantly shorter than in the fed or fasted state 84) ; transit times during fasted and fed dosing regimens were similar in these studies. FG is sensitive to the abundance of drug metabolizing enzymes which could be influenced by genetics and diet. Intestinal epithelial cells express a variety of enzymes involved in phase I and II reactions 85) ; CYP3A 86 90) and UGTs 91) are probably the most influential enzymes affecting gut wall metabolism and hence FG. Certain pathological changes to the GI system, e.g. celiac disease, have been shown to reduce the expression of CYP3A in the gut wall. 92) Inter-individual variation in abundance and regional distribution of intestinal drug metabolizing enzymes 89) can also influence bioavailability of drugs with certain characteristics (e.g. high enzyme affinity, poor gut wall permeability). The formulation attributes (e.g. slow Fig. 4. The inter-individual variation in regional ph values of GI fluids The lines indicate 5 th,25 th,50 th,75 th and 95 th percentile of observed ph values for each segment of the GI tract. These variations may or may not propagate to the bioavailability of drugs depending on whether the release, chemical stability, solubility, dissolution or permeability are ph-dependent or not (key: STO=stomach; DUO= duodenum; PRO=proximal small intestine; MID=mid small intestine; DIS=distal small intestine; CAE=caecum; ASC=ascending colon; TRA=transverse colon; DES=descending colon; R/S=sigmoid colon or rectum; FEC=feces) [adapted from Fallingborg et al ) ; courtesy of Dr Sibylle Neuhoff (Simcyp Limited)] rate of drug release) may also affect gut wall metabolism. 19) Moreover, variation in blood flow can influence gut first pass metabolism of highly permeable drugs, as has been indicated for midazolam in cirrhotic patients. 93,94) Thus, known heterogeneity of blood supply to different segments of the gut, and its variation under pathophysiological conditions (e.g. food intake), 95) should be incorporated in any model that attempts to replicate the human system. Ideally, dynamic changes with time and variable effects in certain sub-groups of the population 80) (Fig. 5) should also be considered. These complexities are another reminder of the important interplay between the system variables and drug attributes in determining inter-individual differences in oral drug absorption. Active secretion of the drug from the gut wall into the gut lumen by the multidrug efflux pump, P-glycoprotein (P-gp), as well as influx and efflux by other transporters may be subject to inter-individual variations affecting transporter abundance and/or activity. However, knowledge in this area is sparse. Heterogeneous distribution along the gut, increasing from proximal to distal small intestine, 96,97) has been suggested for P-gp but this has not been confirmed uniformly in all studies. 98) Thus, these might be specific to each transporter or even each individual. The final element that determines variability in bioavailability is the first pass metabolism of the drugs by the liver (i.e. FH). Variability in FH is directly related to variation in the activity of drug metabolizing enzymes in

11 Virtual Human Populations in Assessing Covariates of ADME 63 Fig. 5. Some sources of variability in oral drug absorption which are associated with food intake Parts (A) and (B) show the blood flow supply to the GI tract in the human proximal jejunum and ileum, respectively. Regional variations in blood flow may influence gut wall metabolism to a variable degree depending on the affinity of the drug to gut wall enzymes, its permeability, and the luminal location of drug release from its formulation. Food intake can increase these flows by up to 30% in addition to its effects on gastric ph which are shown in Part (C). However, it should be noted that the effects of food are transient, which should be considered when incorporating the effects into clinical trial simulations. It is also of interest to note that whilst changes in stomach ph are not age dependent, the dynamics of the return to baseline are different between old and young individuals (Part (C)). The data for Part A and B are taken from DeSesso and Jacobson ) andthedataforpart(c)aretakenfromrussellet al ) the liver, where genetic and environmental factors both play significant roles. Information on enzyme abundance and its variation has been gathered recently 99) ; however this may require regular updating with the advent of high throughput LC-MS assays of proteins ) Some other aspects of variability in metabolism which influence bioavailability (e.g. inter-ethnic differences, co-variation between the enzyme abundance, age, the presence of diseases, concomitant drugs) are discussed in the following sections (under metabolic clearance). However, it should be kept in mind that these are as important in determining variation of absorption as they are in determining the clearance. Sources of Variability in Distribution: The parameters determining inter-individual variability in drug distribution have been outlined in earlier sections. Binding of a drug to plasma proteins and erythrocytes is one factor that influences distribution. However, despite large variation in the binding affinity of different drugs, inter-individual variability tends to be less than that in other pharmacokinetic parameters under nonpathophysiological conditions. 104) The fraction of drug which is unbound in blood is determined by the affinity of the drug to plasma proteins and red blood cells as well as levels of circulating plasma proteins and hematocrit in the individual. The levels of protein in the plasma vary with age, diseases (e.g. liver cirrhosis), pregnancy, trauma and stress. Similarly, age, sex and environmental factors affect the hematocrit (see Fig. 6 for a summary scheme). Models for tissue distribution such as those developed by Poulin and Theil 22) (with later corrections by Berezhkovskiy 21) ) or those by Rodgers et al. 20,105,106) all rely on our knowledge of tissue composition and its covariates. At present such information for humans is sparse and it appears that sex, age, and in some instances ethnicity, are the only known covariates for the size and composition of some, but not all, tissues. Differences in body fat content, in particular between males and females, may lead to differences in the volume of distribution and subsequent variations in the elimination rate and trough concentrations after multiple doses at steady-state. These differences are observed in routine samples taken for therapeutic monitoring (e.g. clozapine; see Rostami-Hodjegan et al. 55) ) and should not be confused with variable exposure that is determined only by the clearance. For example, a highly lipid soluble drug may distribute into the adipose tissue of a female with a high proportion of body fat and become less available to the eliminating organs; so for a given clearance, the half-life would be longer in female and trough concentrations would be higher. A mechanistic description of partitioning to various tissues (Fig. 7) is essential to help determine simple covariates of distribution such as body size, sex and age. Although it is claimed that the use of physiologicallybased models has essentially been limited by their complexity, 107), increasing computer power and reduced run-

12 64 Masoud JAMEI, et al. Fig. 6. Steps in propagating individual variability in fraction of unbound drug in plasma The scheme above indicates the steps for building inter-individual variability in fraction unbound plasma drug concentration with the knowledge of fraction unbound in a test sample. Fraction unbound in plasma (fu) is often seen as a drug-related parameter. However, this is true only if the plasma protein concentration was not variable; which is not true. Knowing the concentration of proteins in the test sample binding affinity constant can be estimated [Step 1] and used to recalculate fu values in a group of individuals with diverse concentration of plasma proteins [Step 2]. Thus, any in silico model that attempts to incorporate the effects of variable plasma proteins in different individuals (e.g. various pathophysiological conditions that influence the plasma proteins) [Step 3] must separate the protein binding affinity (a purer characteristic of the drug binding than fu) from the `system-related' value of protein concentration. Under such a scheme, known covariates associated with plasma protein concentrations can be mirrored on fu with implications for pharmacokinetic parameters such as CL, F H,and V ss (see Table 1) [Key: AAG=Alpha-Acid Glycoprotein; [P]=concentrations of the protein in plasma, fu=fraction of unbound drug in plasma] Fig. 7. The multi-compartment mammillary model of drug distribution The models such as those suggested by Rodgers et al. 105,106,142) give the ability to separate drug attributes from those of the system and to incorporate inter-individual variability in partitioning of various drugs to different tissues in each individual if the individual tissue compositions are known. [K values are `on' and `off' binding rate constants]. See the text for information on availability of parameter values (or otherwise) in human populations. times have removed this restriction. Nonetheless, the limited availability of information regarding tissue size and composition in certain patient groups and individuals can be considered as a hurdle in using the models for determining clinically relevant covariates of drug distribution. The only exceptions relate to extreme pathophysiological conditions such as morbid obesity, very low hematocrit (transplant patients) and severe cirrhosis