APPLICATION OF A GENERIC PHYSIOLOGICALLY-BASED PHARMACOKINETIC MODEL TO THE ESTIMATION OF XENOBIOTIC LEVELS IN HUMAN PLASMA

Number of references: 32 Number of words in the Abstract: 248 Number of words in the Introduction: 375 Number of words in the Discussion: 1500 Non-standard abbreviations: ADME, absorption, distribution, metabolism and elimination; AUC t1-tlast-DN , dose-normalized AUC from the first to the last recorded time points; CL int , hepatic intrinsic metabolic clearance; fu p , fraction unbound in plasma; fu t , fraction unbound in the interstitial fluid; IQ, interquartile; MLFE, mean log fold error; PBPK, physiologically-based pharmacokinetic; PK, pharmacokinetic(s); QSPR, quantitative structure-property relationship; wMLFE, weighted MLFE. DMD #4838 3 ABSTRACT: Estimation of xenobiotic kinetics in man frequently relies upon extrapolation from experimental data generated in animals. In an accompanying paper, we have presented a unique, generic, physiologically-based pharmacokinetic model, and described its application to the prediction of rat plasma pharmacokinetics from in vitro data alone. Here we demonstrate the application of the same model, parameterized for human physiology, to the estimation of plasma pharmacokinetics in man, and report a comparative evaluation against some recently published predictive methods that involve scaling from in vivo animal data. The model was parameterized through an optimization process, employing a training set of in vivo data taken from the literature, and validated using a separate test set of published in vivo data. On average, the vertical divergence of the predicted plasma concentrations from the observed data, on a semi-log concentration-time plot, was 0.47 log units. For the training set, more than 80% of the predicted values of a standardized measure of AUC were within threefold of the observed values; over 70% of the test set predictions were within the same margin. Furthermore, in terms of predicting human clearance for the test set, the model was found to match or exceed the performance of three published interspecies scaling methods, all of which showed a distinct bias towards overprediction. We conclude that the generic PBPK model, as a means of integrating readilydetermined in vitro and/or in silico data, is potentially a powerful, cost-effective tool for predicting human xenobiotic kinetics in drug discovery and risk assessment. DMD #4838 4 Physiologically-based pharmacokinetic (PBPK) models are mathematical descriptions of the flow of blood throughout the body, developed for the simulation of xenobiotic absorption, distribution and elimination. Such models have been employed by scientists from a number of different disciplines who are interested in the simulation and prediction of exposure (Grass and Sinko, 2002; Leahy, 2003). The application of a generic form of a PBPK model to the prediction of xenobiotic plasma levels in rat following an intravenous dose has been reported in an accompanying publication (Brightman et al., 2005). Here we describe the work that we have done to parameterize the same PBPK model for man, and to assess the reliability of the model in estimating plasma levels of xenobiotics, where these values are known from experimentation. In addition, we draw comparisons with alternative methods for predicting human pharmacokinetic properties that involve extrapolation from experimental data generated in animals. Just as there are numerous published compound-specific PBPK models for the rat that utilize data derived from in vivo studies (Sugita et al., 1982; Igari et al., 1983; Tsuji et al., 1983; Bernareggi and Rowland, 1991; Kawai et al., 1994; Blakey et al., 1997), there are many examples of comparable PBPK models for man that rely upon scaling from in vivo animal data in order to simulate the human pharmacokinetics of a particular compound, and frequently incorporate observed clearance data from human subjects (Igari et al., 1983; Sawada et al., 1985; Tsuji et al., 1985; Bernareggi and Rowland, 1991; Kawai et al., 1994; Kawai et al., 1998). The PBPK model for man presented herein appears to represent the only truly generic model to be published to date, since it has been parameterized for human physiology, independently of any specific compound, and the manner in which in vivo distribution and elimination kinetics are predicted is DMD #4838 5 the same for all xenobiotics. Furthermore, the only compound-dependent inputs that the model requires are readily determined in vitro, or even in silico. In this paper, we have concentrated on predicting the in vivo pharmacokinetics of compounds for which plasma levels have been determined following an intravenous dose. Work that we have done to extend the model to predict both human and rat plasma levels following an oral dose will be reported separately. DMD #4838 6 Methods Model Inputs. A generic PBPK model, which enables the prediction of the pharmacokinetic behaviour of any given compound dosed intravenously in a specified human population, without recourse to data derived through in vivo studies, is presented herein. The compound-dependent inputs required by the model are the same as those listed previously (Brightman et al., 2005). Model Description. The PBPK model is based upon that published by Bernareggi and Rowland (Bernareggi and Rowland, 1991), as shown in their The additional features of the model, including adaptations to facilitate modelling of 'diffusionlimited' distribution of an intravenously administered compound into the various tissues and organs, as well as the various processes involved in renal excretion, are described in detail in the accompanying paper, discussing its application to the prediction of rat in vivo pharmacokinetics (Brightman et al., 2005). Model Parameters. The physiological parameters used in the model were obtained from the literature and are given in the Appendix (which is available online as supplemental data); these were scaled according to the actual body weights of the subjects used in the clinical studies being simulated. Tissue and organ volumes were largely derived from a single comprehensive compilation of physiological data for use in pharmacokinetic models (Brown et al., 1997), supplemented by data from other sources for skin (Mapleson, 1963) and testes (Spector, 1956) DMD #4838 7 volumes, and represent the extravascular (combined interstitial fluid and intracellular space subcompartments) volumes only. Blood flow rates were taken almost exclusively from Bernareggi and Rowland (Bernareggi and Rowland, 1991), with the exception of the blood flow to the testes (Williams and Leggett, 1989). The glomerular filtration rate and urine flow rate were from Davies and Morris (Davies and Morris, 1993) and Tang-Liu et al. (Tang-Liu et al., 1983), respectively, whilst the renal tubular lumen volume was obtained from a textbook of physiology (Pitts, 1974). A haematocrit of 0.441 (Altman and Dittmer, 1971) was assumed. Parameterization of the distribution and elimination components of the generic PBPK model for human required the development of a number of correlation models. Two such models, for the prediction of parameters corresponding to the effective in vivo lipophilicity and plasma protein binding, were derived through a process of optimization of the performance of the PBPK model, as described in the companion paper (Brightman et al., 2005). A comprehensive training set of in vivo data was used for this purpose, and is described in greater detail below. The following parameters were derived as detailed previously (Brightman et al., 2005): permeability-surface area products for organ and tissue distribution; intracellular space/interstitial fluid (unbound) partition coefficients; blood/plasma concentration ratio (R); parameters governing renal excretion; hepatic microsomal intrinsic clearance (CL int ) and fractions unbound in plasma (fu p ) and interstitial fluid (fu t ). Stochastic simulations were performed in order to incorporate known variability in the subject body weights and imprecision in the values of the physicochemical parameters, as described previously (Brightman et al., 2005). for 69 different compounds, and was derived from numerous published clinical studies of intravenous dosing in human subjects. The compounds in this diverse training set were drawn from many therapeutic areas. The required model inputs for the training set compounds were obtained as described previously (Brightman et al., 2005). As before, no attempt was made to eliminate compounds from the training set on the basis of any features of the in vivo pharmacokinetic behaviour. Test Dataset. In order to objectively evaluate the performance of the model, an independent set of in vivo test data was constructed. This consisted of 39 instances of plasma concentration-time data for 18 compounds dosed intravenously in human subjects. These data were derived from the literature. The test set compounds were varied in terms of physicochemical properties, and represented diverse therapeutic areas. The requisite model inputs for these compounds were derived as described previously (Brightman et al., 2005). Calculation of the Plasma Concentration Weighted Mean Log Fold Error (wMLFE). For each pair of in vivo and simulated plasma concentration-time profiles, the log fold prediction error was determined at each simulated time point for which there were corresponding in vivo data, and the mean of these errors over all time points was calculated, to give an overall mean prediction error for each instance of simulated data. The wMLFE represents the weighted mean of these individual means. The weights used in the calculation arise from there being multiple instances and/or sources of in vivo data for several compounds, and hence the contribution of DMD #4838 9 each individual log fold prediction error to the overall mean is weighted accordingly; i.e., so that each compound contributes equally, whatever the number of instances of in vivo data for that compound. DMD #4838 10 Results Model Validation. For any given set of input data, output from the PBPK model is in the form of a predicted plasma concentration-time profile. When stochastic simulations are performed for a single set of input data, each iteration generates a predicted profile, and hence the total output consists of a population of profiles that reflect the inherent uncertainty in the input data. Examples of typical simulation results, plotted on the same axes as the corresponding in vivo data, are given in For the remainder of the results presented here, the median of the population of predicted profiles generated from each set of input data was used as an individual estimate of the plasma concentration time course. In order to assess the overall performance of the model in terms of successfully predicting in vivo plasma levels, the plasma concentration wMLFE was determined for both the training and test sets; this statistic corresponds to the mean vertical deviation (in log units) of a simulated data point from a corresponding observed data point on a semi-log plot of plasma concentration versus time. A plasma concentration wMLFE of 0.47 was determined for both the training set and the test set. By way of illustration, the median predicted profiles shown in The frequency distributions of the actual mean fold errors in plasma concentration prediction for both sets of compounds are shown in DMD #4838 12 Although the primary outputs from the PBPK model are the predicted in vivo plasma concentration-time profiles that are generated, these can be utilized for estimation of standard pharmacokinetic (PK) parameters of interest, including area under the concentration-time curve (AUC), allowing direct comparisons to be made with analogous in vivo data. Since different methods of extrapolating AUC from zero time to the first time point and from the last time point to infinity can vary in the estimates they yield, the simulation results were compared to observed data in terms of a standardized parameter, the dose-normalized AUC from the first to the last recorded time points (AUC t1-tlast-DN ). The capability of the PBPK model to accurately predict AUC t1-tlast-DN in man has been evaluated in terms of the median values and interquartile (IQ) ranges of the predicted/observed ratios, for both the training and test sets, as given in For many of the compounds in the PBPK model training set there were multiple instances of in vivo data, derived from one or more published sources. The statistics shown for the PBPK model DMD #4838 14 have therefore been calculated on the basis of the weighted mean fold prediction error for each compound, and hence are directly comparable with those for the published interspecies scaling methods, which were also determined on a per-compound basis. The weighting system employed was the same as that used in the calculation of the plasma concentration wMLFE. There was considerable overlap in the compounds comprising the training sets for the four different methods. The results of this analysis for the individual methods and their respective training sets, as shown in In order to further investigate the relative capabilities of the four methods, a similar statistical analysis of their predictions of clearance in man for the PBPK model test set compounds was carried out, and is summarized in Predictions of human clearance made by the Caldwell et al. and Ward and Smith methods were calculated from published values of rat in vivo clearance for the test set compounds and evaluated against the corresponding published human clearances. Again, the statistics for the PBPK model were derived from the weighted mean fold prediction error for each compound. DMD #4838 15 The predictive accuracy of each method, and particularly of the Ward and Smith method, is generally poorer for the test set compounds than for those comprising the respective training sets, but still at what might be considered an acceptable level for the early stages of drug discovery (Caldwell et al., 2004; Ward and Smith, 2004). However, the deterioration in overall performance is proportionately less for the PBPK model than for the other methods. Consequently, the expected accuracy of the PBPK model seems to be at least comparable to that of Caldwell et al.'s methods in terms of predicting human clearance for a set of novel compounds. For both methods, the same number of predicted values are within a factor of two of the observed values, and an equal or greater proportion of the PBPK model predictions are within a factor of three or four above or below the observed values, although the average fold error for the PBPK model is rather higher. On the evidence shown, the PBPK model would appear however, to considerably outperform the method of Ward and Smith in predicting human clearance for novel compounds. Moreover, the PBPK modelling approach has the obvious advantage over all three interspecies extrapolation methods that no in vivo data are required in order to predict human pharmacokinetics, with a better or similar degree of accuracy. Consideration of the median values of the predicted/observed ratios for the test set It is notable that all four methods were found to substantially over-predict clearance in man of the test set compound digitoxin, which is also a significant outlier for the prediction of AUC t1-tlast-DN by the PBPK model: the predicted/observed ratios for clearance are 14.5, 9.5, 15.0 and 22.1 for the two Caldwell et al. methods, Ward and Smith's method and the PBPK model, respectively. However, the human clearance of quinine, also in the test set, is similarly poorly predicted by the three interspecies scaling methods (predicted/observed ratios of 9.5, 10.5 and 16.6 for Caldwell et al.'s and Ward and Smith's methods, respectively), but more successfully predicted by the PBPK model, with a predicted/observed ratio of 4.3. Finally, the same analysis was applied to a further method published by Ward and Smith for predicting human clearance, requiring scaling from clearance in monkey (Ward and Smith, 2004). The average fold error for the training set data of these authors was found to be 2.04, with 64%, 78% and 89% of the predictions being less than two-, three-and fourfold in error, respectively. The overall level of performance therefore does not greatly exceed that demonstrated by the Caldwell et al. methods when applied to the corresponding training set. Hence, it would have been interesting to establish whether the same deterioration in predictive capability might be shown by this alternative method of Ward and Smith, when applied to the PBPK model test set, as observed for their analogous method that depends on scaling from rat clearance. However, appropriate monkey in vivo data were unfortunately not available for this particular set of compounds. Similarly, Obach et al. have reported several methods of predicting clearance in man (as well as volume of distribution and half-life) from preclinical DMD #4838 17 pharmacokinetic data and/or in vitro data (Obach et al., 1997), but the compounds used in their analysis were not identified, and hence we were unable to directly compare the performances of the PBPK model and these particular methods. DMD #4838 18 Discussion In this and the preceding paper (Brightman et al., 2005), we have demonstrated that a generic PBPK model can produce reliable predictions of mammalian PK following intravenous dosing, for a wide range of organic compounds, from a small number of readily-determined, compounddependent inputs. We expect the overall physiological description to be common to all mammalian species, and that the fundamental model could be successfully adapted to any species by changing the physiological parameters. Since a PBPK model is physiologically realistic, with explicit equations for flows into the major tissues and organs of the body, the compound-dependent parameters required to simulate PK are real properties that can be measured. Consequently, such models can be used to define a minimum set of absorption, distribution, metabolism and elimination (ADME) properties that must be obtained by in vitro or in silico screening in order to predict in vivo PK for any compound. In the same way that clinical trials are the point at which all the experimental data generated during drug discovery and preclinical development are integrated to produce a clear picture of the clinical potential of a new drug, so PBPK models can create the framework for integrating ADME, toxicity and efficacy data throughout discovery and even preclinical development. Our aim is to develop models that are driven by inputs that can be determined within imposed cost and time constraints, and can reliably inform compound selection, by being able to predict human PK with sufficient confidence for a particular phase of the discovery/development process. The PBPK model that we have described in this paper generates predictions of plasma concentrations and clearance in man that appear to be sufficiently reliable to inform compound DMD #4838 19 selection during early drug discovery. We have not, so far, been able to validate the human version of the model against drug-discovery compounds to the same extent as the rat version. Nonetheless, there is no reason to believe that the relative robustness shown by the latter model in its ability to predict PK for compounds that are distant from the training set is not also a property of the human model. We have, however, demonstrated the equal or even superior ability of the current version of the human model to predict clearance when compared to interspecies extrapolation methods. Further development of the model, for example, through incorporation of additional physiological or biochemical processes currently not simulated, extension of the training set and/or the use of additional inputs, will serve to further enhance the predictive capability of the model above methods that are dependent on in vivo animal experimentation. The maximum value that can be extracted from the use of this model within drug discovery largely depends upon the modus operandi of an individual company, drug-discovery programme or project. Significant factors that determine the relative merits of applying the model in any particular situation include: the number of compounds passing through the successive stages of the discovery process; the role of early ADME determination; access to low-cost ADME screens; the methods used for lead expansion and the usage of in silico techniques. Assays are available, for each of the required model inputs, that have sufficiently high throughput to enable human PK prediction using measured values within the timescale dictated by successive synthesis rounds of a typical project in lead optimization. Recent, well-documented changes in practice within the pharmaceutical industry have led to ADME data being generated more thoroughly and earlier in drug discovery than was previously the case. Hence, experimental measurement of some or all of the required inputs will be routinely available during lead optimization for many projects. Alternatively, in silico methods can be used to predict one or more inputs, in order to reduce cost, DMD #4838 20 increase throughput and/or reduce the requirement for the physical compound, the availability of which may be restricted during early discovery. Consequently, we anticipate that the use of rapid in vitro/in silico screening of ADME and physicochemical properties, coupled to the prediction of human PK through PBPK modelling, will enable the optimization of human PK to be a significant determinant of the lead expansion and optimization phases, rendering rat and mouse PK studies increasingly unnecessary. The logical exte