Abstract
The aim of this study was to assess a physiologically based modeling approach for predicting drug metabolism, tissue distribution, and bioavailability in rat for a structurally diverse set of neutral and moderate-to-strong basic compounds (n = 50). Hepatic blood clearance (CLh) was projected using microsomal data and shown to be well predicted, irrespective of the type of hepatic extraction model (80% within 2-fold). Best predictions of CLh were obtained disregarding both plasma and microsomal protein binding, whereas strong bias was seen using either blood binding only or both plasma and microsomal protein binding. Two mechanistic tissue composition-based equations were evaluated for predicting volume of distribution (Vdss) and tissue-to-plasma partitioning (Ptp). A first approach, which accounted for ionic interactions with acidic phospholipids, resulted in accurate predictions of Vdss (80% within 2-fold). In contrast, a second approach, which disregarded ionic interactions, was a poor predictor of Vdss (60% within 2-fold). The first approach also yielded accurate predictions of Ptp in muscle, heart, and kidney (80% within 3-fold), whereas in lung, liver, and brain, predictions ranged from 47% to 62% within 3-fold. Using the second approach, Ptp prediction accuracy in muscle, heart, and kidney was on average 70% within 3-fold, and ranged from 24% to 54% in all other tissues. Combining all methods for predicting Vdss and CLh resulted in accurate predictions of the in vivo half-life (70% within 2-fold). Oral bioavailability was well predicted using CLh data and Gastroplus Software (80% within 2-fold). These results illustrate that physiologically based prediction tools can provide accurate predictions of rat pharmacokinetics.
Obtaining rapid information regarding the pharmacokinetics (PK) of new drug candidates can be a bottleneck in early drug discovery. Considerable resources are required to assess the PK properties of potential drug candidates in vivo in animals. To optimize the use of such in vivo testing, there has been a growing interest in predicting the PK behavior of drug candidates as early as possible (Norris et al., 2000; van de Waterbeemd and Gifford, 2003). If sufficiently reliable, such simulations could also help to select the best candidates for development and to reject those with a low probability of success.
The characterization of a drug's PK requires elucidation of each of the coincident processes of absorption, distribution, metabolism, and elimination (ADME). A large number of methodologies have been established for this purpose, including empirical and physiologically based methods (Boxenbaum and Ronfeld, 1983; Theil et al., 2003). Until recently, PK prediction has been predominantly descriptive, using empirical methods. Although in some cases these methods give good predictions, their physiological basis is low and inaccurate results can be obtained, in particular when there are large interspecies differences in metabolic clearance (Lave et al., 1995; Zuegge et al., 2001; Shiran et al., 2006). To improve on these predictive approaches, it is necessary to explore more mechanistic methods, such as physiologically based pharmacokinetic (PBPK) models (Theil et al., 2003). PBPK models are derived from the knowledge of the underlying physiology of the species and the behavior of drugs within this system. As a result, PBPK models are generic and can be applied to a wide array of structurally diverse compounds (Poulin and Theil, 2002b; Nestorov, 2003). Physiologically based models can be applied for estimation of single in vivo PK parameters (e.g., metabolic clearance), or all data obtained on each ADME process can be interconnected in one global model (i.e., whole body PBPK model), thereby offering the opportunity to provide a priori simulations of complete tissue and plasma concentration-time profiles (Theil et al., 2003).
So far, the use of whole body PBPK models in the pharmaceutical industry has been limited, mainly because of the labor-intensive input required (e.g., tissue-to-plasma partition coefficients, Ptp). During early drug discovery, limited information on compound-specific ADME input parameters and processes can be made available. Therefore, at this early stage, simulations ideally rely on minimal data input. In this context, a variety of mechanism-based ADME prediction tools have been developed for estimation of specific ADME parameters. These prediction tools are built on commonly generated in silico and in vitro compound-specific biochemical and physicochemical properties; therefore, they can be applied in early drug discovery before any in vivo study. For example, prediction tools have been developed for estimation of oral absorption (Agoram et al., 2001; Willmann et al., 2004), Ptp values (Poulin and Theil, 2000; Poulin et al., 2001; Rodgers et al., 2005a), volume of distribution at steady state (Vdss) (Poulin and Theil, 2002a; Rodgers and Rowland, 2006), and metabolic clearance (Houston and Carlile, 1997; Austin et al., 2002; Ito and Houston, 2004). Although these tools may greatly extend the possibility of assessing ADME properties in early drug discovery, in general, their utility and limitations have still been poorly investigated. Furthermore, these prediction tools have been validated mainly on marketed drugs, which tend to have different physicochemical properties compared with those of current early discovery projects (Wenlock et al., 2003).
The primary aim of the present work was to evaluate the recently described mechanism-based prediction tools for their ability to predict individual ADME processes in the rat on a structurally diverse set of neutral and moderate-to-strong basic compounds. Modeling the individual PK processes offered the advantage that the prediction accuracy as well as its sensitivity toward physicochemical diversity could be addressed independently for each PK parameter. Moreover, in our efforts to apply these methods in drug discovery mode, the validity of various approaches toward disregarding binding factors as well as the use of computed parameters were assessed. Although prediction of PK in rat is obviously not of ultimate interest for the pharmaceutical industry, a validated prediction strategy in rat can then be scaled to human and was shown to improve prediction of human PK (Jones et al., 2006). To be able to do this, we used a large body of data generated over a time frame of several years. The compounds in the data set cover a broad range of small molecules from a variety of discovery programs, presenting a great challenge since each method must be applicable across a broad range of structural chemotypes and therapeutic areas.
Materials and Methods
Sources of in Vitro, in Silico and in Vivo Parameters. The compounds used in the analysis were taken from those brought into early development at Johnson & Johnson Pharmaceutical Research and Development (Beerse, Belgium). Compounds (n = 50) were selected based on the availability of historical data on experimentally derived in vitro data: fraction unbound in plasma (fup), dissociation constant (pKa), n-octanol:water partition coefficient of the non-ionized species (logPow), aqueous solubility (Sw) at defined pH, in vitro half-life in liver microsomes (in vitro t1/2), and blood-to-plasma concentration ratio (RB) (Table 1). All experimental values of fup were obtained using DIANORM equilibrium dialysis (Pacifici and Viani, 1992). Solubility data were obtained at 20°C after overnight incubation of drug powder in water or dosing vehicle at defined pH conditions between 4.0 and 7.5. All structure-based predictions of logPow,pKa, and jejunal permeability were obtained using ADMETPredictor version 1.3.2 (Simulations Plus Inc., Lancaster, CA).
In vivo pharmacokinetic data of the 50 compounds studied in rat are summarized in Table 2. Plasma concentration versus time profiles were obtained from at least three male rats for each route of administration. Test compound was given intravenously by bolus injection through a catheter implanted into the jugular vein and orally by gavage. Noncompartmental analysis was performed using WinNonLin version 4.01 (Pharsight, Mountain View, CA) to calculate the total blood clearance (CLtot) from the relationship CLtot = Dose/AUC, and Vdss was determined as Vdss = Dose · AUMC/(AUC)2. Absolute oral bioavailability (F) was calculated as the ratio of dose-normalized AUC after oral and intravenous administration using the mean of individual AUCs. Experimental Ptp values have been determined under in vivo conditions (single oral or intravenous dose) as the ratio of the AUC (calculated over a minimum of five time points), assuming pseudo-equilibrium.
Prediction of Hepatic Blood Clearance (CLh) from in Vitrot1/2 Data. The in vitro t1/2 of each compound was determined in pooled male rat liver microsomes (protein concentration 1 mg/ml, substrate concentration 5 μM) using a typical screening assay as described previously (Kantharaj et al., 2003). Conversion of in vitro t1/2 data (minutes) to the intrinsic clearance (CLint, ml/min/kg) was performed as follows: The scale-up factors from protein to grams of liver and from grams of liver to kilograms body weight (kg b.wt.) are 45 mg/g liver (Houston, 1994) and 40 g liver/kg b.wt. (Naritomi et al., 2001), respectively. Conversion of CLint to CLh involved the use of either the well stirred model (eq. 2) or the parallel tube model (eq. 3) (Wilkinson, 1987): where Qh is the hepatic blood flow (55.2 ml/min/kg) (Davies and Morris, 1993) and fum is the fraction unbound in microsomes. Both models were examined using three different variations: 1) assuming that no binding parameter has an impact on clearance (fup/RB = fum = 1); 2) incorporating binding to blood constituents only (fum = 1); and 3) incorporating both blood and microsomal binding. For microsomal protein binding, fum in the incubation mixture at 1 mg/ml microsomal protein was calculated by eq. 4 (Austin et al., 2002):
Calculation of CLh from in Vivo Data. After oral dosing, CLh values were determined from eq. 5 by use of the apparent oral blood clearance (CLoral) values (Iwatsubo et al., 1997). For all calculations, renal clearance (CLR) was considered to be negligible and the fraction of dose that has reached the portal vein (FDp) was taken to be 1. For compounds JNJ5, JNJ15, JNJ16, JNJ27, and JNJ28, where only CLtot was available, CLh was assumed to be equal to CLtot.
Prediction ofPtp andVdss According to Rodgers et al. (2005a) (Method Vd1). A detailed derivation of the equations for predicting the Ptp values of moderate-to-strong bases is provided in the article by Rodgers et al. (2005a). In brief, it is assumed that electrostatic interactions of moderate-to-strong bases with acidic phospholipids is the primary factor controlling the distribution of these drugs within the body. In addition, moderate-to-strong bases also dissolve in tissue water, and the un-ionized form can partition into neutral lipids and neutral phospholipids, so that Ptp of unbound drug (Ptpu) can be calculated using the following equation: where V is the fractional tissue volume content of extracellular water (EW), intracellular water (IW), neutral lipids (NL), and neutral phospholipids (NP). [AP]t is concentration of acidic phospholipids in tissue. Values for pHp (pH of the plasma) and pHIW were taken to be 7.4, and 7.0, respectively. pKa represents the dissociation constant of the monoprotic base (cutoff in this study, pKa ≥ 6.8). For this study, P was taken to be the vegetable oil:water partition coefficient of the non-ionized species (Pvow), which was calculated from logPow, as follows (Leo et al., 1971):
Ka is the association constant of the compound with the acidic phospholipids, and is calculated from eqs. 8 to 10: where BC is the red blood cell, E:P the erythrocyte-to-plasma concentration ratio, and Ht the hematocrit content in blood (assumed to be 45%). Values for pHp,pHIW, and pHBC were taken to be 7.4, 7.0, and 7.22, respectively. In eq. 8, P was taken to be the Pvow. It is important to note that the use of Pow in eq. 8 yielded negative KaBC values (and thus negative Ptpu) for compounds having both a low RB value and a high Pow value. This problem was overcome by using Pvow instead of Pow for calculation of KaBC (eq. 8) as well as for calculation of Ptpu of both adipose and nonadipose tissues (eq. 6).
For calculation of Vdss, each predicted Ptpu value for bone, brain, intestine, heart, kidney, liver, lung, muscle, skin, spleen, and adipose tissue was multiplied by fup to obtain the corresponding Ptp value for each tissue. Vdss equals the plasma volume (Vp) in addition to the sum of each Ptp multiplied with its respective tissue volume (Sawada et al., 1984): where V is the fractional body volume in l/kg tissue (t), erythrocyte (e), and plasma (p). All tissue-specific input information on Vt, Ve, and Vp was obtained from the literature (Rodgers et al., 2005a). It is important to note that the use of Pvow in both eqs. 6 and 8 was not found to affect the overall Vdss prediction accuracy compared with Rodgers et al. (2005a), who have used Pow and Dvow (i.e., vegetable oil:buffer partition coefficient of both the non-ionized and ionized species at pH 7.4) as input parameters for calculation of nonadipose and adipose Ptpu, respectively (data not shown).
Prediction ofPtp andVdss According to Poulin and Theil (Method Vd2) (Poulin and Theil, 2000;Poulin et al., 2001). A detailed derivation of the tissue composition-based equations for predicting the Ptp in adipose and nonadipose tissue is provided by Poulin and Theil (Poulin and Theil, 2000; Poulin et al., 2001). In brief, it is assumed that a drug distributes homogeneously into each tissue (and plasma) by passive diffusion. Consequently, the drug partitions between lipids and water and binds reversibly to common proteins present in plasma and tissue interstitial space: where V is fractional tissue (t) or plasma (p) volume content of neutral lipids (NL), phospholipids (PH), and water (W). The data on rat tissue volumes was obtained from the literature (Poulin and Theil, 2000; Poulin et al., 2001). To obtain Dvow, logPvow was first estimated from data on logPow using eq. 7, and logPvow was subsequently converted to logDvow using the Henderson-Hasselbalch equations (Poulin and Theil, 2002a). The values of the fraction unbound in tissue (fut) were estimated from data on fup, as follows: where RA is the ratio of albumin concentration found in tissue over plasma. For adipose tissue, RA was set to 0, whereas for nonadipose tissue, RA was set to 0.50 (Poulin and Theil, 2002a).
For calculation of Vdss, Ptp values were determined in bone, brain, intestine, heart, kidney, liver, lung, muscle, skin, spleen, and adipose using eqs. 12 and 13, and subsequently inserted into eq. 11 as described under Method Vd1.
Prediction of in Vivo Half-Life (in Vivot1/2). Both method Vd1 and method Vd2 for predicting Vdss were combined with the predicted hepatic plasma clearance values (CLh,plasma = CLh · RB) from both the well stirred (Fig. 1, A–C) and parallel tube models (Fig. 1, D–F) to generate predictions of in vivo t1/2 using the following formula:
Prediction ofF. The first method (method F1) used for predicting F was based on eq. 16 and only accounts for hepatic first pass metabolism and thus accounts neither for potential limitations on absorption nor potential first pass metabolism by the gut (i.e., FDp = 1). CLh values were obtained using the well stirred model, assuming that fum = fup/RB (See also Fig. 1A).
In the second method (method F2), Gastroplus Simulation (version 5.1.0; Simulations Plus Inc.) was used. For all simulations, Gastroplus was provided with experimentally derived data (Table 1) on logPow,pKa, Sw at defined pH, or solubility in dosing vehicle at defined pH, dose administered (Table 2), and first pass metabolism. First pass metabolism was assumed to occur only in the liver and was calculated from the CLh as described under method F1. As no experimental data were available on permeability estimates, in silico estimates of human jejunal permeability (Table 1) were obtained by the artificial neural network model in ADMETPredictor (version 1.3.2; Simulations Plus Inc.) and subsequently converted to rat permeability by the built-in correlation of Gastroplus. The variation of solubility and permeability within different regions of the gut was assumed to depend on the pH of that region and was calculated from the pKa and logD values of the drug (Gastroplus Opt logD model).
Success Criteria. Success of predictions was assessed by the root mean squared prediction error (rmse) and the average-fold error (afe) as measures of precision and bias, respectively, with equal value to under- and overpredictions:
A prediction method with an afe ≤2 was considered successful. Predicted CLh and Vdss values were deemed accurate if they agreed with mean experimental in vivo values within a factor of 2 (Obach, 1999; Poulin and Theil, 2002a). Ptp predictions were considered successful if they agreed with experimental values within a factor of 3, as has been suggested previously (Poulin and Theil, 2000; Rodgers et al., 2005a).
Results
Physicochemical and Pharmacokinetic Properties of the Compounds. The molecular weight ranged from 287 to 777. The lipophilicity was high, with logPow ranging between 1.1 and 5.5. The majority of the compounds (n = 36) were defined as bases with a basic pKa ≥ 6.8, and the remainder (n = 14) were defined as neutrals with a basic pKa < 6.8. The fup ranged from 0.001 to 0.82. Solubility was highly variable with values ranging from <0.001 mg/ml to 192 mg/ml. Although this study was performed retrospectively, the physicochemical properties of our data set in Table 1 have remained representative of compounds currently being considered as preclinical development candidates (Table 1, JNJ37, JNJ38, JNJ44, JNJ45). The majority of compounds were moderate-to-strong bases, characterized by high lipophilicity and low free fraction in plasma.
Prediction of CLh in Rat. The objective of the experiments described herein is to test more exhaustively the possibility of predicting CLh by disregarding the fraction unbound in both blood (fup/RB) and microsomes (fum) in eqs. 2 and 3. The validity of the assumption that binding to blood constituents in vivo would be similar to binding to microsomal protein in vitro (i.e., fup/RB is cancelled out by fum) would greatly extend the use of hepatic clearance models in early drug discovery, where data on fup, RB, and fum are usually not available. CLh was predicted using both the well stirred (eq. 2) and parallel tube (eq. 3) models under three variations: 1) disregarding all binding values (fup/RB = 1 and fum = 1), 2) including only blood binding (fum = 1), and 3) including both blood and in vitro microsomal binding. Graphical comparisons of the prediction methods are shown in Fig. 1 (A–F), and the parameters for the accuracy of predictions are summarized in Table 3. In general, disregarding all binding in either model of hepatic extraction (Fig. 1, A and D) yielded the best agreement between observed and predicted CLh values (approximately 85% within 2-fold of observed). Using only the blood binding value (Fig. 1, B and E) in either model of extraction yielded very poor predictions, with a strong bias toward underprediction (approximately 20% within 2-fold of observed). Including both binding factors improved significantly the accuracy of the predictions (approximately 60% within 2-fold of observed) compared with blood binding only; however, many of the underpredictions remained (Fig. 1, C and F). The latter finding suggests that both binding factors would not necessarily cancel out. Poor predictions were obtained for compounds with an in vitro t1/2 more than 60 min (n = 13), with less than 50% of predicted values within 2-fold of observed (Table 3).
Prediction ofVdss: Method Vd1 versus Vd2. There were 47 compounds that had adequate intravenous PK data for assessment of Vdss predictions (Tables 1 and 2). Since method Vd1 is only to be applied to basic compounds, the utility and accuracy of both methods was first compared solely on all moderate-to-strong bases (basic pKa ≥ 6.8, n = 35). Figure 2 illustrates the correlations between the observed and predicted values of Vdss using methods Vd1 and Vd2, respectively. Method Vd1 resulted in highly accurate predictions of Vdss with more than 80% within 2-fold of observed (Fig. 2A). In contrast, method Vd2 was a poor predictor of Vdss in this analysis in that only 60% of predictions were within 2-fold of observed (Fig. 2B). For the neutral compounds (n = 12), only method Vd2 was applicable and the results are also shown in Fig. 2B (open circles) and Table 4.
To assess the bias of the predictions, the prediction error (expressed as the log of predicted/observed ratio) was plotted as a function of predicted Vdss for both method Vd1 (Fig. 2C) and method Vd2 (Fig. 2D), and various statistical parameters were calculated (Table 4). Remarkably, method Vd2 yielded a considerable bias with general underpredictions and overpredictions for bases and neutrals, respectively. The underpredictions were predominantly associated with strong bases (pKa > 8.0), characterized by either high lipophilicity (logPow > 4) or low protein binding (fup > 0.5). The majority of the neutral compounds (11 of 12 compounds) showed high lipophilicity (logPow > 3.5).
To explore whether it was possible to reduce the number of experimentally determined input parameters, Vdss predictions (n = 35) using experimentally derived logPow and pKa values were compared with those using their calculated counterparts generated by ADMET-Predictor. The various statistical parameters calculated are given in Table 4. Using computed input, for method Vd1 a considerable increase in bias (afe) and a decrease in accuracy (rmse) were seen compared with experimentally derived logPow and pKa values.
Prediction of Tissue Distribution (Ptp) in Rat: Method Vd1versus Method Vd2. To directly compare methods Vd1 and Vd2 in their accuracy to predict Ptp, only the basic compounds of Table 1 were considered. Tissues considered for correlation analysis were muscle, liver, lung, heart, brain, and kidney because only for these tissues were sufficient data on experimental Ptp available (n ≥ 19). Graphical comparisons of method Vd1 (Fig. 3, closed circles) and Vd2 (Fig. 3, open circles) in each tissue are shown in Fig. 3 (A–F), and the parameters for the accuracy of predictions are summarized in Table 5.
Trends for Ptp predictions were seen similar to those observed for Vdss predictions. In general, method Vd1 yielded more accurate predictions compared with method Vd2, where there was a tendency to underpredict. For both methods, acceptable predictions were obtained in muscle and heart tissue, whereas underpredictions were most pronounced in liver and lung. In contrast, in brain tissue, overpredictions were obtained for both methods; however, the bias was significantly more pronounced using method Vd2.
Estimation of in Vivot1/2 in Rat. Combining both methods for predicting Vdss (methods Vd1 and Vd2) with all those for predicting CLh (Fig. 1) resulted in a total of 12 t1/2 calculation methods. To directly compare methods Vd1 and Vd2, in vivo t1/2 predictions were performed on all moderate-to-strong bases, providing they had an in vitro t1/2 < 60 min (Table 1, n = 25); compounds with an in vitro t1/2 > 60 min were not included because CLh predictions were not found to be accurate (Table 3). The parameters for the accuracy of predictions are summarized in Table 6. Overall, in vivo t1/2 calculations based on CLh predictions that included protein binding were generally more inaccurate than those based on CLh predictions that disregarded all protein binding. Among the latter, the best in vivo t1/2 calculations were based on method Vd1 (Fig. 4), whereas those based on method Vd2 showed similar accuracy but higher bias, due to underpredictions of Vdss (data not shown).
Prediction ofF. Only compounds with experimental solubility data (Table 1) and an in vitro t1/2 < 60 min were considered for bioavailability (F) predictions (n = 30). The graphs of predicted versus observed F are shown in Fig. 5, and the parameters for the accuracy of the predictions are given in Table 7. In the first method (method F1), hepatic first pass metabolism was considered as the sole determinant influencing F, thus assuming that all compounds were 100% absorbed in the portal vein (FDp = 1). Despite its simplicity, this method yielded accurate predictions with more than 80% of predictions within 2-fold of observed (Fig. 5A). These results suggest that for most compounds within this analysis, F was predominantly governed by hepatic first pass metabolism. This hypothesis was supported by the Gastroplus simulation (method F2) depicted in Fig. 5B. The comparable results within both panels of Fig. 5 confirm that for most of the compounds within this analysis, neither permeability nor solubility rate-limited issues significantly affected the absorption process. However, it is important to note that for all simulations of Fig. 5B, the influence of solubility enhancers (cyclodextrin or polyethylene glycol), if any were present in the dosing vehicle, on Gastroplus solubility input data were taken into account (Table 7, Method F2 Simulation 2). When the effect of such additives on solubility was ignored (Table 7, Method F2 Simulation 1), Gastroplus strongly underpredicted F for seven compounds (i.e., JNJ2, JNJ26, JNJ38, JNJ41, JNJ42, JNJ43, and JNJ47), resulting in an increased rmse and a decrease in number of accurate predictions within 2-fold of observed.
Discussion
In recent years, PBPK models and prediction tools based on in vitro and in silico input parameters have become more popular. However, the predictive utility of such tools within drug discovery has been on the whole poorly evaluated, with only a few reports in the literature (Parrott et al., 2005a,b). In general, the prediction accuracy as well as the body of in vitro data needed for prediction of a particular parameter will vary depending on the approach, the type of chemistry, and the prediction system used. Therefore, in our efforts to apply these methods in drug discovery mode, the objectives of this study were: 1) to assess the validity of various approaches toward binding factors when predicting CLh from microsomal data; 2) to perform a comparative evaluation of two recently published tissue composition-based equations for their accuracy and input requirements to predict tissue distribution (Ptp and Vdss); 3) to explore the use of computed physicochemical input and predicted microsomal binding; and 4) to combine the aforementioned parameters and predict in vivo t1/2 and F.
Overall, CLh was a well predicted process within this analysis, irrespective of the type of model of hepatic extraction used (Fig. 1). Because the differences between the two liver models are minimal, we suggest that the most commonly adapted well stirred model could continue to be applied within our in-house chemistry. Poor correlations were obtained for compounds that did not show significant turnover in vitro. Remarkably, literature reports usually do not include such compounds within their analyses, most probably because poor correlations are to be expected (e.g., enzyme saturation, phase 2 metabolism, etc.) Because compounds are often selected for further development based on this property, such data should not be left out from correlation analysis.
The decision whether to incorporate plasma and microsomal protein binding in CLh predictions remains controversial. In drug discovery, one is frequently faced with the situation in which data on fum and fup are not available; therefore, it is an attractive option to assume that both parameters may cancel out. The inclusion of both unbound fractions has been suggested to be the general applicable approach. However, our results and those of others demonstrate that, in the case of some compound classes, especially basic ones, disregarding all binding values may yield the most accurate predictions (Obach, 1997, 1999). Nevertheless, one must recognize that accurate predictions from scaled microsomal data are only possible when the compound is mainly cleared by oxidative microsomal metabolism, and neither extrahepatic metabolism nor renal or biliary clearance significantly contributes to CLtot. For some of the compounds within our analysis, multiple forms of elimination are known to be present (e.g., JNJ3, JNJ6, JNJ22, JNJ41); in these cases, scaled microsomal data should not be able to fully project CLtot.
Mechanistic equations to predict tissue distribution have been developed by Poulin and Theil (Poulin and Theil, 2000, 2002a; Poulin et al., 2001) (method Vd2), who reported that for a set of 123 marketed drugs, 80% of the predicted Vdss values were within 2-fold of observed. In our study, however, the overall prediction accuracy on Vdss and Ptp of these equations was reduced to approximately 60% within 2-fold of observed (Fig. 2). This may be related to the different physiochemical properties in the compound sets. The majority of the compounds examined in this study were moderate-to-strong bases, with a median logPow and fup of 4.03 and 0.038, respectively. In contrast, the data set used by Poulin and Theil (Poulin and Theil, 2000, 2002a; Poulin et al., 2001) consisted mainly of weak acids/bases and neutrals, with a median logPow and fup of 2.17 and 0.25, respectively. Figure 2 reveals that strong bases (pKa > 8.0) characterized by high lipophilicity (logPow > 4) or low-to-moderate lipophilicity combined with low protein binding (logPow < 2 and fup > 0.5) are prone to underprediction. In particular, strong bases with low lipophilicity and low plasma binding were strongly underpredicted (e.g., JNJ3, JNJ8, JNJ22, JNJ34, JNJ36). Some underpredictions might be explained by sensitivity analysis illustrating that Vdss predictions become insensitive to logPow changes above 4 and fup changes below 0.01 (Poulin and Theil, 2002a; Parrott et al., 2005b). In this study, Vdss prediction of all lipophilic strong bases (logPow > 4, pKa > 8) was blunted around 4 to 5 l/kg (e.g., JNJ4, JNJ6, JNJ11, JNJ26, JNJ29, JNJ32, JNJ37). Remarkably, Vdss predictions of highly lipophilic strong bases with fup values of ≥0.05 seemed to be more prone to underprediction (e.g., JNJ4, JNJ29, JNJ32, JNJ37).
The tissue composition-based equation of Rodgers et al. (2005a) (method Vd1) performed significantly better on the prediction of both Vdss and Ptp, irrespective of lipophilicity (Fig. 2). This divergence in prediction accuracy between method Vd1 and Vd2 can be explained as ionic interactions between charged drugs and acidic phospholipids are neglected in method Vd2. Basic drugs tend to have larger Vdss values due to favorable ion-pair interactions of the basic centers with the acidic head groups of the phospholipid membranes (Rodgers et al., 2005b). In the model of method Vd2 it is assumed that macromolecular binding of drugs in tissue interstitial fluids (i.e., fut) is driven by binding to macromolecules similar to those present in plasma; therefore, fut is simply calculated from fup data (eq. 14). However, in the case of strong bases (pKa > 8.0), the calculated fut value will most likely underestimate tissue binding. In particular, for bases that have low protein binding, the calculated fut can be expected to deviate more from actual. Overall, these results suggest that the model of Rodgers et al. (2005b) should be used for moderate-to-strong bases, and certainly for our in-house chemistry. However, the major disadvantage is the prerequisite for an additional input factor (i.e., RB). To reduce the required input, some efforts were devoted to explore computed parameters. Unfortunately, a decrease in accuracy was seen using computed parameters compared with experimentally derived ones (Table 4).
A similar trend between methods Vd1 and Vd2 was demonstrated for prediction of data on Ptp (Fig. 3). With the exception of muscle and heart tissue, a poor accuracy and strong bias of method Vd2 was demonstrated. Bias resulted predominantly from pronounced underprediction. Despite the overall better performance of method Vd1, underpredictions still prevailed in lung and liver tissues. One possible explanation is lysosomal trapping, since these acidic organelles are abundant in lung and liver, and extensive distribution to these tissues has been reported for many basic drugs (MacIntyre and Cutler, 1988; Ishizaki et al., 1998). Importantly, significant lysosomal trapping may affect the first passage of a basic compound through the lung; thus, the estimation of Vdss in vivo from plasma concentration-time profiles may be biased. In liver tissue, potential clearance processes, enterohepatic circulation, or active uptake processes may also have influenced experimental Ptp values. In brain tissue, overpredictions were predominantly apparent for both methods. The main assumption with these equations is that passive diffusion governs tissue distribution. Brain penetration is, however, more restrictive and selective, and several active transport mechanisms are likely to play a key role. Inaccurate predictions will also prevail if in vivo Ptp values are estimated under non-steady-state conditions, for example, after a single oral or intravenous dose. Nevertheless, Ptp values based on the AUC values should theoretically equate to those obtained at steady state (Rodgers et al., 2005a).
In combining CLh and Vdss predictions to predict in vivo t1/2, all methods were combined to remain unbiased (Table 6). As expected, most successful combinations in predicting in vivo t1/2 consisted of those methods that were most successful in predicting the independent parameters. An important consideration in the prediction of in vivo t1/2 is the fact that t1/2 values may exhibit multiphase plasma concentration versus time profiles; therefore, terminal t1/2 is most likely to be underpredicted when using CLh and Vdss predictions. However, in drug discovery it is not of tremendous importance to target prediction of in vivo t1/2 as absolute values but rather as an ability to place compounds in appropriate dosing regimens.
Predictions of F were generally successful, even when only hepatic metabolism was considered (Fig. 5A). It is acknowledged that such methods should be treated carefully since, by definition, they undervalue the potential impact of absorption and first pass metabolism limitations by the gut mucosa. More advanced PBPK models have now become available to predict FDp and F. For example, the model underlying Gastroplus Software is known as the Advanced Compartmental Absorption and Transit Model (ACAT model) (Agoram et al., 2001). Gastroplus takes drug-specific input data (solubility, permeability) as well as metabolic first pass estimates and uses these to predict FDp and F. In this study, Gastroplus simulation of F yielded results similar to those from predictions based solely on CLh data (Fig. 5), suggesting that hepatic first pass metabolism was a major determinant influencing F. Interestingly, accuracy of Gastroplus simulations seemed to improve when the influence of solubility enhancers, if present in the dosing vehicle, were taken into account. A similar issue was found by Parrott et al. (2005b), who corrected for increased solubility that would occur in the presence of bile salts.
In summary, we have compared various mechanism-based ADME prediction tools to predict CLh, tissue distribution (Vdss and Ptp), in vivo t1/2, and F using a diverse data set covering many targets and chemotypes. Although such prediction tools are said to be generic and thus applicable to a wide array of structurally diverse compounds, our results and those obtained by others demonstrate that diverse physicochemical properties as well as invalid model assumptions may put compounds beyond the validity of the approach. Therefore, it is suggested that the utility of each prediction tool should be validated on a class-by-class basis. Physiologically based prediction of human PK with an initial validation in animals is currently under evaluation in our department.
Footnotes
-
Article, publication date, and citation information can be found at http://dmd.aspetjournals.org.
-
doi:10.1124/dmd.106.014027.
-
ABBREVIATIONS: PK, pharmacokinetic(s); ADME, absorption, distribution, metabolism, and elimination; AUC, area under the curve; AUMC, area under the first moment curve; CLh, hepatic blood clearance; CLh,plasma, hepatic plasma clearance; CLint, intrinsic clearance; CLoral, oral blood clearance; CLR, renal clearance; CLtot, total blood clearance; E:P, erythrocyte-to-plasma concentration ratio; F, absolute oral bioavailability; FDp, fraction of dose that has reached the portal vein; fum, fraction unbound in microsomes; fup, fraction unbound in plasma; Ht, hematocrit content in blood; PBPK, physiologically based pharmacokinetics; Dvow, vegetable oil:buffer partition coefficient of both the non-ionized and ionized species at pH 7.4; Pow, n-octanol:water partition coefficient of the non-ionized species; Ptp, tissue-to-plasma partition coefficient; Pvow, vegetable oil:water partition coefficient of the non-ionized species; Qh, hepatic blood flow; RB, blood-to-plasma concentration ratio; Sw, aqueous solubility; Vdss, volume of distribution at steady state.
- Received November 22, 2006.
- Accepted January 24, 2007.
- The American Society for Pharmacology and Experimental Therapeutics