Comparative Assessment of In Vitro In Vivo Extrapolation Methods used for Predicting Hepatic Metabolic Clearance of Drugs

Transcription

1 Comparative Assessment of In Vitro In Vivo Extrapolation Methods used for Predicting Hepatic Metabolic Clearance of Drugs PATRICK POULIN, 1 CORNELIS E. C. A. HOP, 2 QUYNH HO, 2 JASON S. HALLADAY, 2 SAMI HADDAD, 3 JANE R. KENNY 2 1 Consultant, 4009 Sylvia Daoust, Québec City, Québec G1X 0A6, Canada 2 DMPK, Genentech Inc., South San Francisco, California Département de Santé Environnementale et Santé au Travail, IRSPUM, Faculté demédecine, Université demontréal, Montréal, Québec H3T 1J4, Canada Received 23 May 2012; revised 26 June 2012; accepted 17 July 2012 Published online 13 August 2012 in Wiley Online Library (wileyonlinelibrary.com). DOI /jps ABSTRACT: The purpose of this study was to perform a comparative analysis of various in vitro--in vivo extrapolation (IVIVE) methods used for predicting hepatic metabolic clearance (CL) of drugs on the basis of intrinsic CL data determined in microsomes. Five IVIVE methods were evaluated: the conventional and conventional bias-corrected methods using the unbound fraction in plasma (fu p ), the Berezhkovskiy method in which the fu p is adjusted for drug ionization, the Poulin et al. method using the unbound fraction in liver (fu liver ), and the direct scaling method, which does not consider any binding corrections. We investigated the effects of the following scenarios on the prediction of CL: the use of preclinical or human datasets, the extent of plasma protein binding, the magnitude of CL in vivo, and the extent of drug disposition based on biopharmaceutics drug disposition classification system (BDDCS) categorization. A large and diverse dataset of 139 compounds was collected, including those from the literature and in house from Genentech. The results of this study confirm that the Poulin et al. method is robust and showed the greatest accuracy as compared with the other IVIVE methods in the majority of prediction scenarios studied here. The difference across the prediction methods is most pronounced for (a) albumin-bound drugs, (b) highly bound drugs, and (c) low CL drugs. Predictions of CL showed relevant interspecies differences for BDDCS class 2 compounds; the direct scaling method showed the greatest predictivity for these compounds, particularly for a reduced dataset in rat that have unexpectedly high CL in vivo. This result is a reflection of the direct scaling method s natural tendency to overpredict the true metabolic CL. Overall, this study should facilitate the use of IVIVE correlation methods in physiologically based pharmacokinetics (PBPK) model Wiley Periodicals, Inc. and the American Pharmacists Association J Pharm Sci 101: , 2012 Keywords: disposition; microsomes; clearance; unbound fraction; computational ADME; in vitro in vivo extrapolation; in vitro in vivo correlation; IVIVE; pharmacokinetics; PBPK modeling Abbreviations used: AAG, alpha1-acid glycoprotein; AFE, average-fold error; AL, albumin; BCS, bioclassification system; BDDCS, biopharmaceutics drug disposition classification system; CCC, concordance correlation coefficient; CL, clearance; CL int,intrinsic CL; fu inc, unbound fraction in incubation; fu liver, unbound fraction in liver; fu p, unbound fraction in plasma; fu p-app, apparent unbound fraction in plasma; IVIVE, in vitro in vivo extrapolation; K m, Michaelis Menten constant; PhRMA, Pharmaceutical Research and Manufacturers of America; PLR, plasma-to-whole liver concentration ratio of albumin; Q liver, blood flow rate to liver; R BP, blood-to-plasma concentration ratio; RMSE, root-mean-squared error. Correspondence to: Dr. Patrick Poulin (Telephone: ; patrick-poulin@videotron.ca) INTRODUCTION Various methods are available to predict human pharmacokinetics with some based on preclinical in vivo data and others utilizing human in vitro data. in vitro methods are most convenient because they require minimal amount of compound and do not necessitate animal studies. However, the predictivity of in vitro methods depends on the model used and the input Journal of Pharmaceutical Sciences, Vol. 101, (2012) 2012 Wiley Periodicals, Inc. and the American Pharmacists Association 4308 JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012

2 COMPARATIVE ANALYSIS OF IVIVE METHODS 4309 parameters. Factors influencing the predictive performance of an in vitro in vivo extrapolation (IVIVE) method for hepatic metabolic clearance (CL) are related to several input parameters; namely, binding terms such as the unbound fraction in plasma (fu p ) and in incubation medium (fu inc )aswellastheintrinsic CL (CL int ) and liver blood flow rate (Q liver ) Wan et al. 1 studied the impact of these input parameters on the CL estimate in rat and human datasets. The authors concluded that the simplified IVIVE method, disregarding binding data (i.e., direct scaling), might be sufficiently good for IVIVE evaluations. Obach 2 also suggested disregarding all binding data to predict human CL for basic and neutral compounds, whereas for acidic compounds, he suggested including all binding terms (i.e., fu p /fu inc ). Recently, Berezhkovskiy et al. 12,13 and Poulin et al., 14 both of whom also studied the impact of the binding terms on CL estimates, presented two novel IVIVE methods. Berezhkovskiy s method consists of replacing fu p with an apparent fu p (fu p-app ) that considers drug ionization differences between the plasma and liver cells. Poulin et al. 14 further analyzed the concept of binding terms and suggested converting the value of fu p-app to an unbound fraction in the liver (fu liver ) to take also into account the role of extracellular binding proteins on the passive uptake of drugs in hepatocytes. Using a dataset of 25 drugs, the Poulin et al. 14 method showed the greatest accuracy as compared with other IVIVE methods on the basis of several statistical parameters. 14 Recently, Halifax and Houston 15 used a larger dataset and demonstrated superior precision and lower bias in the majority of cases for the novel method of Poulin et al.; however, these authors are not in total agreement on the mechanistic justification of the method advocated by Poulin et al. 14 Instead, Halifax and Houston 15 proposed an empirical scaling method involving a conventional model, but corrected for the average-fold error (AFE) (i.e., the conventional bias-corrected method). Therefore, a consensus on the use of IVIVE methods could not be agreed upon, and hence, further testing is needed. The purpose of this study was to further investigate the published IVIVE methods by using large and diverse datasets from human, monkey, dog, and rat. This study might help to identify potential outlier drugs and apply further refined IVIVE methods to identify the strengths and limitations of these methods. METHODS The overall strategy consisted of evaluating the effect of the following scenarios on predictive performance of various IVIVE methods for CL based on microsomal data: (a) the preclinical and human datasets, (b) the extent of plasma protein binding [i.e., drugs bound to albumin (AL), drugs bound to alpha1-acid glycoprotein (AAG), and drugs highly bound in plasma], (c) the magnitude of CL under in vivo conditions (i.e., very low, low, medium, and high CL), and (d) the extent of drug disposition based on the biopharmaceutics drug disposition classification system (BDDCS) and/ or bioclassification system (BCS) categorizations. 16 Furthermore, we explored the effect of hepatic uptake on CL estimations by using the current IVIVE methods for a reduced dataset of drugs in rats. For this dataset, the Poulin et al. 14 method was compared with the direct scaling method. We theorized that the direct scaling method may be advantageous when CL in vivo is unexpectedly high because this method naturally overpredicts the true metabolic CL, as is reported in the literature. 1,2,7,13 Finally, we present a sensitivity analysis to demonstrate how the different IVIVE methods vary with input parameters related to drug ionization, plasma protein binding, and/or CL int. Comparative Analysis of IVIVE Methods Five IVIVE methods that have undergone previous comparative assessments were the focus of further evaluation in this study. 1,2,14,15 These IVIVE methods are (a) the conventional and conventional bias-corrected methods using the fu p,(b)the Berezhkovskiy method in which the fu p is adjusted for drug ionization on either side of the plasma membrane on the basis on ph differences, (c) the Poulin et al. 14 method using the fu liver to adjust in addition for protein-facilitated uptake because of the potential ionic interactions between the plasma-proteinbound-drug complex and the cell surface of the hepatocytes, and (d) the direct scaling method that does not consider any binding corrections. Table 1 summarizes all equations related to these IVIVE methods. Recently, Halifax and Houston 15 reported an empirical method, the conventional bias-corrected method, which involves multiplying the predicted CL values from the conventional method with the corresponding average bias of underprediction to reduce the underprediction. The average bias of underprediction was obtained from the AFE observed for each dataset (i.e., each prediction scenario) of this study. This empirical method was also evaluated in this study. The wellstirred model was considered for the purpose of this study. Furthermore, the parallel tube model was also used for high CL compounds for all IVIVE methods tested because it is expected that the prediction accuracy for these drugs will increase with the parallel tube model. 2 Estimation of the Input Parameters The five IVIVE methods scale CL int determined in microsomes from in vitro-to-in vivo conditions by using a physiologically based scaling factor based on DOI /jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012

3 4310 POULIN ET AL. Table 1. IVIVE Methods Tested in this Study as Reported from the Literature 1,2,12 14 Methods Conventional method Adapted from Berezhkovskiy a Poulin et al. 14 method Direct scaling method Model Equations of Plasma CL CL = Q liver R BP CL int,invivo fu p Q liver R BP + CL int,invivo fu p (1) CL = Q liver R BP CL int,invivo fu p app/fu inc Q liver R BP + CL int,invivo fu p app/fu inc (2) where fu p app = and F 1 = F 1 fu p 1+(F 1 1)fu p (3) f unionized plasma f unionized liver,cells (4) and f unionized neutral = 1 (5) f unionized monoprotic acid = 1/ [ 1 + ( 10 ph pka )] (6) f unionized monoprotic base = 1/ [ 1 + ( 10 pka ph)] (7) f unionized diprotic bases = 1/ [ 1 + ( 10 pk a1 ph + 10 pk a1+pk a2 2pH )] (8) CL = Q liver R BP CL int,invivo fu liver /fu inc Q liver R BP + CL int,invivo fu liver /fu inc (9) where PLRfu fu liver = p app 1+(PLR 1)fu p app (10) CL = Q liver R BP CL int,invivo Q liver R BP + CL int,invivo (11) AL, albumin; AAG, alpha1-acid glycoprotein; Q liver, liver blood flow rate; R BP, blood-to-plasma ratio; fu p, unbound fraction in plasma; fu p-app, apparent unbound fraction considering the ph gradient; fu liver, unbound fraction in liver considering the protein-facilitated uptake and ph gradient; fu inc, unbound fraction in incubation medium (microsomes); F I, ionization factor; f unionized, fraction unionized; CL int, intrinsic clearance; PLR, plasma-to-whole liver ratio of AL or AAG. For the conventional-bias corrected method, the AFE value is used as described in the Method. a The calculation of fu p-app has been made from the binding isotherm to cover both the highly and the less bound drugs, and to avoid fu p-app > 1 for basic drugs. hepatic microsomal recovery from whole liver to convert CL int in vitro in :L/(min mg proteins) to ml/(min kg body weight) for CL int in vivo (i.e., 900, 800, 1908, and 1720 mg protein per kg body weight in humans, monkeys, dogs, and rats, respectively). 1,2,6,14 In addition, Q liver was taken to be 20.7, 43.6, 30.3, and 65 ml/(min kg) in humans, monkeys, dogs, and rats, respectively. 6,14 In in vitro experiments it is doubtful whether the ph gradient between the extracellular and intracellular spaces of liver is maintained. Because microsomes were incubated with a buffer at ph 7.4, the in vitro measurements of CL int do not account for ph differences in extracellular and intracellular spaces of liver. Thus, application of the methods of Berezhkovskiy 12,13 and Poulin et al. 14 may represent a significant improvement because the ph gradient for the ionized drugs is taken into account (Table 1). The ph values used were 7.4 for plasma and 7.0 for liver cells. The ph value of liver cells was chosen to represent the mean of most of the values reported in the literature, which ranged from 6.89 to As mentioned, the Poulin et al. 14 method consists of adjusting the well-stirred model for the proteinfacilitated uptake in addition to the effect of the ph gradient. These effects are considered in the estimation of fu liver (Table 1). The traditional assumption is that equilibrium between the free and protein-bound drug is instantaneous, such that the metabolism process is driven by a constant supply of unbound drug concentration in plasma. However, in this study, we also have assumed that unbound drug concentration in liver and plasma differs because of the proteinfacilitated uptake of drugs into hepatocytes, and hence, the current fu liver represents a correction of the observed fu p to take into account the extracellular protein binding in plasma relative to liver in addition to the effect of the ph gradient. Thus, the current fu liver should not be totally comparable to the overall unbound fraction determined in liver homogenates, for example, which would consider binding to diverse components (e.g., lipids). It was determined that AL plays a major role in the protein-facilitated uptake of drugs in hepatocytes as compared with the AAG. 14 Consequently, Poulin et al. 14 taken into account the extracellular protein binding in plasma relative to liver [or the plasma-towhole liver concentration ratio (PLR) of AL or AAG] in the calculation of fu liver. The experimentally determined value of PLR for AL in rat and human is about ,20,21 For monkey, the same PLR value of 13.3 was used in this study. Briefly, the PLR value of AL was obtained by converting the measured plasma-toextracellular fluid concentration ratio into plasma whole liver ratio to be in agreement with the wellstirred model, which assumes an homogeneous drug distribution in the liver. 14 The volume of interstitial fluid in liver is used as the conversion factor because the level of AL in liver cells ( 0.4 mg/g) is negligible as compared with the extracellular fluid, and hence, of serum ( 40 mg/g). 14,22 However, for dog, the volume of interstitial fluid in the liver is far greater JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012 DOI /jps

4 COMPARATIVE ANALYSIS OF IVIVE METHODS 4311 than in other species ( 0.29 vs. 0.18). 23,24 Consequently, when species difference in the volume of interstitial space is considered, the PLR value in dog is estimated to be 8.5 (i.e., assuming that the plasma-toextracellular fluid concentration ratio of AL for liver is species invariant). For each drug bound to AAG, the PLR value for this protein was set to unity, and hence, fu liver is equal to fu p-app 14 (Table 1). The values of PLR can be debated and were questioned by Halifax and Houston. 15 However, it should be emphasized that the values of PLR across species were chosen in a truly prospective fashion considering the measured PLR data presented in the literature without knowing the predictivity of this IVIVE method. Datasets The drug datasets and the corresponding experimentally determined input parameters are presented in the Appendix (see Supplementary information). The datasets consist of drugs obtained from the literature 1 7,14,25 and drug candidates synthesized at Genentech (South San Francisco, California). A Pharmaceutical Research and Manufacturers of America (PhRMA) initiative 6,7 also published concise and complete datasets with experimental data for several drugs in rat, dog, monkey, and/or human. For the studies of Wan et al. 1 and Hosea et al., 5 only the neutral drugs were considered because the essential pk a values of the proprietary compounds were not presented in these studies. In addition, the in vitro and in vivo CL as well as the essential input parameters was experimentally determined in rat for 21 compounds at Genentech as described in the Appendix (see Supplementary information). As mentioned, a reduced dataset for rat containing five drugs for which CL in vivo is governed by an active uptake process (digoxin, ketanserin, PhRMA #37, #40, and #43) is also presented in the Appendix (see Supplementary information). 6,26 28 The total number of drugs studied here is 139, covering a large range of drug properties. The in vitro CLdatausedinthis study were obtained from in vitro metabolic assays by using plasma-free microsomal incubations. The main binding protein (AL or AAG) was identified for each literature drug on the basis of mechanistic studies published in the literature. 14 When this information was not available (28 proprietary compounds), binding to AAG was assumed to be preferred for strongly basic drugs, whereas binding to AL was assumed to be preferred for acidic and neutral drugs. However, for each Genentech compound, the main binding protein was determined experimentally as described in the Appendix (see Supplementary information) (i.e., the binding ratio of human AAG to human AL was determined by using a dialysis system). Modeling Assumptions The drugs used in this study are thought to be eliminated by hepatic oxidative metabolic CL under in vivo conditions in each species. The liver metabolism for several of the marketed compounds is governed mainly by cytochrome P450 (CYP) enzymes. 1 6,14 Although this information was unknown for the propriety compounds, we assumed that their metabolism was governed similarly. Transporter-mediated processes that could possibly be responsible for drug uptake or drug efflux from hepatocytes were neglected, except for the five drugs in the reduced rat dataset. Because the microsomal data were determined in plasma-free incubations, it was assumed that drug distribution from plasma to hepatocytes was not impeded by limited diffusion processes under in vivo conditions. In the metabolic stability assays, the substrate concentration was expected to be well below the apparent Michaelis Menten constant (K m ). It was also assumed that the in vivo CL of drugs follows the free drug hypothesis according to the well-stirred model. In this case, binding to plasma proteins was assumed to be reversible and not saturated at the conditions studied. Evaluation of Predictive Performance The prediction accuracy was assessed by comparing predicted versus observed values of CL by using several statistical parameters. 6,15,24 The AFE and rootmean-squared error (RMSE) were calculated and are presented for each prediction method. Furthermore, the concordance coefficient of correlation (CCC) global is presented, which evaluates the global degree to which pairs of predicted and observed data fall on the line of unity passing through the origin. Specificfold errors of deviation between the predicted and observed values (percentage of fold error 2and 5) were also calculated. Finally, plots of predicted versus observed CL values are also presented for each IVIVE method. Sensitivity Analysis Theoretical simulations of plasma CL in human over a large range of CL int were investigated with the Poulin et al., 14 Berezhkovskiy, and direct scaling IVIVE methods. A span of predicted CL values is presented for two drug examples, namely, a strong acid (pk a = 2) and a base (pk a = 10) either bound (fu p = 0.01) or less bound (fu p = 0.9) in plasma. For the Poulin et al. 14 method the acidic and basic drugs were defined to bind to AL and AAG, respectively. DOI /jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012

5 4312 POULIN ET AL. RESULTS Comparative Assessment for Various IVIVE Methods for Predicting CL Several IVIVE calculation methods of CL were compared by using the same drug datasets, and the comparative assessment was made on the basis of several statistical parameters. The overall statistical summary in terms of accuracy, precision, and correlation is listed in Tables 2 4 for the different scenarios of prediction. The plots of predicted versus observed CL values for each method are shown in Figures 1 5, whereas Figure 6 compares the precision bias across the IVIVE methods. The sensitivity analysis is presented in Figure 7. All Datasets The results of this study confirm that the Poulin et al. 14 method is robust and showed the greatest accuracy among the IVIVE methods tested in the majority of the prediction scenarios. The statistical analysis is generally in favor of the Poulin et al. 14 method because superior, or at least comparable, statistics were obtained by this method as compared with those obtained by other methods (Tables 2 4); this is also corroborated graphically in Figures 1 5. The empirical, conventional bias-corrected method suggested by Halifax and Houston 15 showed lower predictivity as compared with the other methods studied here. This empirical method was, therefore, not fully investigated in this study. Moreover, the conventional bias-corrected method required analysis of several datasets to first determinate the AFE correction factor, which depends on the composition of the datasets (i.e., number and disposition of compounds) as shown by the variability in the AFE values across the various prediction scenarios of this study. Therefore, it is not surprising to observe that the resulting AFE values of the conventional bias-corrected method are close to unity in most cases, however, other statistical parameters are not much improved over other methods (Tables 2 4). An important observation is that the Poulin et al. 14 method generally provides AFE values close to unity, whereas the Berezhkovskiy and conventional methods provide lower AFE value and the direct scaling method gives the highest AFE value. Therefore, a systematic underprediction of CL in vivo was observed for the Berezhkovskiy and conventional methods, whereas the direct scaling method overpredicted the CL as compared with the Poulin et al. 14 method. In particular, the Poulin et al. 14 method presented the lowest bias in precision especially for the following prediction scenarios: AL-bound drugs and those drugs with low fu p and low CL values (Fig. 6). Accordingly, the global CCC value is closest to unity for the Poulin et al. 14 method, and hence, the RMSE Table 2. Comparative Assessment of IVIVE Methods Used to Predict Plasma CL According to Preclinical and Human Datasets Percentage of Twofold or Less Error Percentage of Fivefold or Less Error AFE RMSE CCC All datasets (n = 134) Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling Human (n = 48) Poulin et al Berezhkovskiy Conventional Direct scaling Dog and monkey (n = 22) Poulin et al Berezhkovskiy Conventional direct scaling Rat (this study) (n = 21) Poulin et al Berezhkovskiy Conventional Direct scaling Rat (literature) (n = 43) Poulin et al Berezhkovskiy Conventional Direct scaling AFE, average-fold error; RMSE, root-mean-square error; CCC, concordance correlation coefficient. JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012 DOI /jps

6 COMPARATIVE ANALYSIS OF IVIVE METHODS 4313 Table 3. Comparative Assessment of IVIVE Methods Used to Predict Plasma CL According to Drugs Bound to AL and AAG and Drugs Highly Bound in Plasma (fu p 0.01) Percentage of Twofold or Less Error Percentage of Fivefold or Less Error AFE RMSE CCC Highly bound (n = 10) Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling AL bound (n = 72) Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling AAG bound (n = 62) Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling AFE, average-fold error; RMSE, root-mean-square error; CCC, concordance correlation coefficient. Table 4. Comparative Assessment of IVIVE Methods Used to Predict Plasma CL According to the Magnitude of CL In Vivo Percentage of Twofold or Less Error Percentage of Fivefold or Less Error AFE RMSE CCC Very low CL ( 5% of Q liver )(n = 13) Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling Low CL ( 25% of Q liver )(n = 44) Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling Medium CL (n = 70) Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling High CL ( 75% of Q liver )(n = 20) Well-stirred model Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling Parallel tube model Poulin et al Berezhkovskiy Conventional Conventional bias corrected Direct scaling AFE, average-fold error; RMSE, root-mean-square error; CCC, concordance correlation coefficient. DOI /jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012

7 4314 POULIN ET AL. Poulin et al. Conventional Observed CL Predicted CL Figure 1. Predicted CL versus observed CL for the IVIVE method proposed by Poulin et al. 14 according to all datasets (CCC = 0.91 and n = 134). The solid line indicates the best fit (unity). Short and long dashed lines on either side of unity represent twofold and threefold error, respectively. CL is in ml/(min kg). Berezhkovskiy Predicted CL Observed CL Observed CL Figure 3. Predicted CL versus observed CL for the conventional IVIVE method according to all datasets (CCC = 0.55 and n = 134). The solid line indicates the best fit (unity). Short and long dashed lines on either side of unity represent twofold and threefold error, respectively. CL is in ml/(min kg). Conventional bias corrected Observed CL Predicted CL Figure 2. Predicted CL versus observed CL for the IVIVE method proposed by Berezhkovskiy according to all datasets (CCC = 0.79 and n = 134). The solid line indicates the best fit (unity). Short and long dashed lines on either side of unity represent twofold and threefold error, respectively. CL is in ml/(min kg). value is the lowest. Furthermore, this method often shows the greatest accuracy based on the fold errors of deviation between the predicted and observed values (percentage of fold error 2and 5). It is useful to take a closer look at the prediction scenarios in more detail because more information can be obtained on the predictivity of one method compared with another. Predictivity of Preclinical and Human Datasets On the basis of all statistical parameters, the Poulin et al. 14 method was the best performing prediction method when considering the current human dataset (n = 48; Table 2). This method resulted in 88% of predicted CL within twofold error as compared with the observed values in human, whereas this number was lower for the other methods tested (25% 69%). Again, no systematic underestimation or overestimation of CL in human was observed with the Poulin Predicted CL Figure 4. Predicted CL versus observed CL for the conventional bias-corrected IVIVE method according to all datasets (CCC = 0.75 and n = 134). The solid line indicates the best fit (unity). Short and long dashed lines on either side of unity represent twofold and threefold error, respectively. CL is in ml/(min kg). Observed CL Direct scaling Predicted CL Figure 5. Predicted CL versus observed CL for the direct scaling IVIVE method according to all datasets (CCC = 0.74 and n = 134). The solid line indicates the best fit (unity). Short and long dashed lines on either side of unity represent twofold and threefold error, respectively. CL is in ml/(min kg). JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012 DOI /jps

8 COMPARATIVE ANALYSIS OF IVIVE METHODS 4315 AL-bound Low f up Prediction scenarious Low CL Figure 6. AFE values obtained from the different IVIVE methods for different prediction scenarios. Black squares, direct scaling method; red circles, Poulin et al. 14 method; green triangles, Berezhkovskiy method; and pink diamonds, conventional method. et al. 14 method (AFE = 0.96), whereas the other methods underpredicted or overpredicted human CL (AFE = 0.21, 0.52, and 1.66). Similar findings are obtained for the dog and monkey datasets. For the rat datasets, the predictivity of the Poulin et al. 14 method is superior (n = 21; Genentech dataset) or comparable (n = 43; literature dataset) to the predictivity tested by other IVIVE methods. In addition, the AFE values resulting from the predictivity for the two rat datasets is still closest to unity for the Poulin et al. 14 method (Table 2). Furthermore, it is observed in Table 2 that the Poulin et al. 14 method is the most accurate for the rat dataset of Genentech compounds for which all input parameters have been experimentally determined (including the main binding protein). Table 2 indicates that the predictive performance decreases in the rat datasets as compared with the human dataset for the conventional, Berezhkovskiy, and Poulin et al. 14 IVIVE methods. The average CL of the rat datasets represents 62% of the liver blood flow rate, whereas this number decreases to 32% for the human dataset (Appendix; see Supplementary information). This means that the compounds for which we have human data may have less complex pharmacokinetics in humans as compared with the rats (i.e., mainly CYP-mediated oxidative metabolism). This aspect is considered further in the Discussion section. Conversely, the predictive performance of the direct scaling method increased in the rat datasets as compared with the human dataset. Therefore, the direct scaling method was further evaluated with a reduced rat dataset of five drugs for which their CL in vivo is unexpectedly high (Table A2; see Supplementary information) as well as according to the effect of the BDDCS and/or BCS categorization of the literature drugs 16 (Table A1; see Supplementary information) as presented below. Calculated plasma CL [ml/(min/ kg)] a Acidic drug;f up = b Acidic drug;f = 0.9 up c Basic drug;f up = 0.01 d Basic drug;f up = Intrinsic CL [ml/(min/ kg)] Figure 7. Theoretical simulations of plasma CL in human over a large range of CL int.(a)an acidic drug highly bound to AL (pk a = 2 and fu p = 0.01). (b) An acidic drug not significantly bound to AL (pk a = 2 but fu p = 0.9). (c) A basic drug highly bound to AAG (pk a = 10 and fu p = 0.01). (d) A basic drug not significantly bound to AAG (pk a = 10 but fu p = 0.9). The other input parameters in the IVIVE methods were set equal to 1 (R BP ) and 0.5 (fu inc ). Some simulations are similar and, therefore, overlap (b and d). Poulin et al. s 14 method is shown by a black solid line, Berezhkovskiy s method is shown by a dashed gray line (long dashed lines), and the direct scaling is shown by a dashed blue line (short dashed lines). DOI /jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012

9 4316 POULIN ET AL. Assessing Prediction of CL in Rat for a Reduced Dataset The Poulin et al. 14 method correctly predicted rat CL within threefold error for only one of the five drugs in which CL in vivo in rat was unexpectedly high because of active uptake, whereas the direct scaling method correctly predicted three drugs. The direct scaling method provided improved prediction performance in this circumstance because it naturally shows trends of overprediction of CL as compared with the Poulin et al. 14 method. The reason being the resulting AFE value is much greater for the direct scaling method (0.43) than for the Poulin et al. 14 method (0.08) for these five drugs. Predictivity According to the BDDCS and/or BCS Class Relevant interspecies differences were observed in predictivity between class 1 and 2 compounds when either the BDDCS or BCS information was used. For example, for the rat dataset, the Poulin et al. 14 method generated 88% (15/17) of the CL estimates within twofold error for class 1 compounds, but this number decreased significantly to 30% (3/10) for class 2 compounds. Conversely, the predictivity for the human dataset was more comparable between the two drug classes, with 89% (25/28) and 79% (11/14) of the CL estimates within twofold error for class 1 and 2 compounds, respectively. Another example is that the prediction of CL in vivo of ketanserin and ritanserin (BDDCS class 2) in human is within twofold error, whereas in the rat, the predicted values are greatly underpredicted (up to a factor of sixfold) (not shown). The disparity between predictions of class 2 compounds in rat is particularly evident with the Poulin et al. 14 method as compared with the direct scaling method. The reason is that the direct scaling method provides higher predicted CL values, which is benifical when CL in vivo is higher than expected in rat especially for those class 2 compounds. Predictivity According to Plasma Protein Binding As expected, the Poulin et al. 14 method showed superior predictive performance particularly for drugs bound mainly to AL and those highly bound in plasma (fu p values 0.01) (Table 3). For example, the Poulin et al. 14 method provided AFE values ranging from 1.06 to 1.21, whereas the conventional and Berezhkovskiy methods obtained much lower AFE values ( ). The direct scaling method did not perform better as AFE values ranged from 1.72 to 2.56 (Fig. 6). Note that the Berezhkovskiy and Poulin et al. 14 methods behave the same for (basic) compounds bound to AAG because PLR is equal to unity for AAG as explained in the Methods section (i.e., fu liver is equal to fu p-app ) (Tables 1 and 3). JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012 Predictivity According to the Magnitude of Drug CL An important difference in predictivity was observed among the IVIVE methods tested when the magnitude of CL in vivo of the drug was considered (Table 4). For drugs with a low CL in vivo ( 25% of Q liver ), the Poulin et al. 14 method is by far the most accurate prediction method. This is even more evident for drugs with a very low CL in vivo ( 5% of Q liver ). Indeed, 62% of CL predictions for drugs with low and very low CL in vivo are within twofold error with the Poulin et al. 14 method, whereas the estimates range from 8% to 31% for the other IVIVE methods. The AFE value for the Poulin et al. 14 method is 1.15, whereas this value is much greater (5.67) for the direct scaling method and is lower for the conventional (0.11) and Berezhkovskiy (0.16) methods (Fig. 6). Conversely, for the medium high CL range (>25% of Q liver ), the Poulin et al. 14 and direct scaling methods are comparable, followed by the conventional and Berezhkovskiy methods. The reasonable predictivity of the direct scaling method particularly for the high CL compounds is in line with previous findings on trends of over-prediction of CL compared to the other IVIVE methods. The results for high CL compounds slightly improved for all IVIVE methods when the parallel-tube model is used as compared with the well-stirred model. This is reflected in the values of AFE and CCC, which are closer to unity, whereas the values of RMSE decreased (Table 4). Sensitivity Analysis The sensitivity analysis demonstrates the importance of compound properties and binding parameters that are reflective of specific mechanistic determinants relevant to prediction of the CL values. Figure 7 illustrates differences among the methods tested, particularly at lower CL int values; at higher CL int values, the IVIVE methods are more similar. Furthermore, the class of drug affected CL predictions. This is more noticeable when an acidic drug is highly bound to AL than when a basic drug is highly bound to AAG. The differences observed in the sensitivity analysis are in accordance with the superior predictivity of the Poulin et al. 14 method especially in the problematic areas of high protein binding and low CL, as reported previously (Tables 3 and 4). Thus, the importance of fu p as an input parameter is obvious, but it should be highlighted that it is experimentally hard to determine fu p accurately for highly bound drugs. DISCUSSION We conducted a comparative analysis of five promising IVIVE methods by using a dataset of 139 drugs obtained in preclinical species and human. The findings of this assessment suggest that the method proposed DOI /jps

10 COMPARATIVE ANALYSIS OF IVIVE METHODS 4317 by Poulin et al. 14 offers a significant improvement in the prediction of CL as compared with other IVIVE methods. In other words, the results confirm the validity of the novel method published by Poulin et al. 14 for calculation of hepatic metabolic CL, which accounts for ph differences in extracellular and intracellular water of liver as well as the protein-facilitated uptake. Berezhkovskiy et al. 13 previously studied the effect of a ph gradient on the estimation of CL int under in vitro conditions, but this study and two published comparative analyses 14,15 demonstrate that considering solely a ph gradient in the IVIVE of CL is not sufficient, and hence, the protein-facilitated uptake should also be included for more accurate extrapolations. Notably, the general success of the novel IVIVE method of Poulin et al. 14 is probably because of the adjustment of the well-stirred model with fu liver. Because no consideration of mechanisms related to the main binding proteins in liver was made with the other IVIVE methods tested, this may explain why they generally provided a lower prediction performance as compared with the Poulin et al. 14 method. In summary, the difference across these methods is most pronounced for (a) AL-bound drugs, (b) low CL drugs, and (c) highly bound drugs (Figs. 6 and 7). Halifax and Houston 15 suggested that the prediction bias was CL dependent and protein binding dependent for all methods, indicating important sources of bias from in vitro methodology. The extent of CL and protein binding are acknowledged to be potential sources of error that could cause discrepancy between in vitro and in vivo values. In this study, relatively lower predictivities were obtained for low CL compounds as compared with high CL compounds as well as for highly bound drugs as compared with less bound drugs (Tables 3 and 4). Halifax and Houston 15 suggested that prediction from microsomes, and particularly from hepatocytes, might be improved beyond any of the methods assessed in this study through the use of an empirical correction factor to eliminate both the average bias and the CL dependency bias. Sohlenius-Sternbeck et al. 29 also presented a method of removing the systematic bias through application of empirical correction factors derived from regression analyses applied to the in vitro and in vivo data for a defined set of reference compounds. However, Sohlenius-Sternbeck et al. 29 optimized their regression equations from hepatocyte data only, and hence, they cannot be applied in the present study where microsomal data were used. Indeed, these two empirical methods are fully dependent on the analysis of the in vivo datasets. In contrast, the Poulin et al. 14 method only requires compound-specific input and no prior analysis of a large dataset to provide CL predictions. The bias for CL and protein binding was reduced in this study with the mechanistic method of Poulin et al. 14 as compared with the empirical technique from Halifax and Houston 15 on the basis of several statistical parameters for problematic areas (i.e., compounds with low fu p and low CL values). Both this study and the literature 14 demonstrate that considering fu liver in IVIVE represents a significant statistical improvement in these problematic areas as compared with other methods (Fig. 6); therefore, justification for the mechanistic modelling proposed here, including that for highly bound drugs, is well supported. The Poulin et al. 14 method is sensitive to the value of PLR, but the value was determined on the basis of the measured data available in the literature, and the analysis presented here provides justification for the value and the novel approach. Mechanistic plasma protein binding assays are required to identify the major binding protein to apply the IVIVE calculation method of Poulin et al. 14 prospectively. In this context, the present study proposed a new experimental setting to define whether AL or AGG is the main binding protein in plasma for 21 Genentech compounds (i.e., when AAG/AL is dialysed against each other, a ratio greater than 0.6 suggested AAG is preferred; Appendix; see Supplementary information). Many highly binding drugs may bind to both proteins (AL or AAG). We assumed either one or the other protein, but in reality the binding ratio of 0.6 is an arbitrary cutoff. For compounds that bind equally to both proteins, we may potentially consider a binding to AL invoking PLR correction in the estimation of fu liver and we will explore that further in a subsequent analysis. Any significant errors in experimental assessment of fu p would confound the predictability of IVIVE, particularly for highly bound drugs. 10,14 For several propriety compounds (i.e., Pfizer, J&J, and PhRMA compounds), the main binding protein was unknown, and so it was estimated on the basis of the class of drug, which may have affected the comparative assessment. The paucity and variability of literature data relating to IVIVE demonstrate the importance of generating experimentally consistent data for the purpose of method validation because the data quality is crucial for the prediction accuracy, particularly for compounds with low fu p and CL values. 1,10,14,30 In light of the fact that discrepancies are observed in CL int,fu inc,andfu p values as well as in in vivo plasma concentrations between laboratories, Beaumont et al. 10 strongly recommend using the same concentration (preferably a low one) of compounds for both the in vitro metabolic stability and binding assays to avoid possible concentrationdependent effects. In fact, the substrate concentration should be well below the apparent K m.another source of error is the variability in the lipid content and composition of liver microsomes, 31 both of which may influence the measure of K m particularly in a human population. DOI /jps JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012

11 4318 POULIN ET AL. Despite potential experimental errors, proteinfacilitated uptake has been reported as an important factor affecting the CL of highly bound drugs under in vitro conditions. Recently, Wattanachai et al. 32 showed that supplementation of human liver microsome incubations with bovine serum AL resulted in a 3.6-fold increase in the microsomal CL int for paclitaxel 6"-hydroxylation, due mainly to a reduction in the K m. A lower K m value is in accordance with a proteinfacilitated mechanism because of the presence of additional ionic interactions between the protein drug complex and microsomes compared with when the drug is used alone in the incubation system. This observation from Wattanachai et al., 32 on the basis of microsomal data, corroborates the findings of several other authors that used hepatocytes data. 14,33 37 The reason is that liver microsomes and membranes of hepatocytes have a similar composition of lipids to which the protein drug complex may potentially aggregate under in vitro and in vivo conditions. 14,31,38 Overall, this is in accordance with the pharmacokinetics in liver of AL-bound drugs 14,33,39 and charged macromolecules (e.g., antibodies), 40,41 which is governed by the binding to the protein and the presence of positive charge on the protein, and, hence, is in line with the use of the current fu liver in IVIVE. Conversely, Halifax and Houston 15 seem not in total accordance with the implication of the protein-facilitated mechanism but these authors did not provide any robust explanation to support their disagreement. Nevertheless, Halifax et al. 42 studied the impact of drug permeability in hepatocytes to understand the poor in vitro-to-in vivo correlation of CL of highly bound drugs. These authors demonstrated that the prediction accuracy was not dependent on the relative permeability of drugs in hepatocytes indicating the absence of a general rate limitation by passive hepatocyte uptake on metabolic CL. Protein-facilitated uptake has been reported to be more important for AL-bound drugs than AAGbound drugs, 14,33 37,39 which may explain why Poulin et al. s 14 method shows improved predictive performance for drugs that bind mainly to AL as compared with the other methods studied. Inversely, drugs that bind mainly to AAG seem to require no additional correction, which may explain why the predictive performance for AAG-bound drugs is more comparable across the IVIVE methods tested (particularly the Berezhkovskiy and Poulin et al. 14 methods) 14 (Table 3). These observations are reflected in the sensitivity analysis depicted in Figure 7. Prediction accuracy has been shown to be generally better in human than in rat (Table 2), probably because it is expected that pharmacokinetics in humans is less complex (e.g., hepatic CL mediated by CYP-oxidative metabolism) as a consequence of the screening processes in drug discovery and develop- JOURNAL OF PHARMACEUTICAL SCIENCES, VOL. 101, NO. 11, NOVEMBER 2012 ment. Here, we point out that the preclinical datasets (particularly the rat datasets) might contain a significant amount of drugs for which CL in vivo is unexpectedly high; this could introduce bias in the comparative analysis. Indeed, the rat datasets present a greater percentage of compounds with high CL as compared with the human dataset (Appendix; see Supplementary information). Moreover, Table 2 shows that in the rat dataset obtained from the literature the relative predictivity of all methods is comparable, as anticipated for high CL drugs as illustrated in Table 4. The rate of transporter-mediated uptake and efflux determines the rate of hepatobiliary elimination, and considerable species variation in the mechanisms of biliary excretion is seen in the literature. 26,43 45 Transporters are, therefore, important determinants of CL in the body Furthermore, a molecular weight threshold was observed for biliary excretion, which is different in rats and humans, particularly for anionic compounds. 44 Consequently, the literature reports lower predictivity of CL for drugs that potentially undergo biliary excretion. 45 Accordingly, this present study reports a difficulty in predicting drug CL in rat of five drugs that potentially undergo biliary excretion. Lam and Benet 27 also demonstrated the effect of hepatic uptake processes on the CL estimate of digoxin using microsomal and hepatocytes data. The direct scaling method has an advantage when CL in vivo of a drug is unexpectedly high; however, careful attention should be given when the direct scaling method is used because this method naturally trends toward overestimation of the true metabolic CL (AFE value > 1), as proven in this study and the literature. 1,2,7,14 Biopharmaceutics drug disposition classification system and/or BCS classification may help to understand the interspecies difference in the prediction accuracy. Overall, only 5% 10% of drugs differ in classification between the BCS and BDDCS. However, for class 1 drugs, the difference in classification between BCS and BDDCS is estimated to occur for about 40% of drugs. 16 Nevertheless, the present study observed a relevant difference in the prediction accuracy between class 1 and 2 compounds, classified according to either the BDDCS or BCS, and this difference is much more important in the rat dataset as compared with the human dataset. Umehara and Camenisch 46 also reported differences in CL estimations in rats across the BDDCS classes. Thus, it appears that there are interspecies differences in the mechanisms of CL under in vivo conditions, which may have influenced the current comparative assessment. Alternatively, one should consider that the BCS/BDDCS classification is based on human data and it is conceivable that classification may change if it were based on rat data. BDDCS and BCS class 2 compounds are expected to be governed by DOI /jps