Prediction of in vivo drug–drug interactions from in vitro data: impact of incorporating parallel pathways of drug elimination and inhibitor absorption rate constant
Abstract
Aims
Success of the quantitative prediction of drug–drug interactions via inhibition of CYP-mediated metabolism from the inhibitor concentration at the enzyme active site ([I]) and the in vitro inhibition constant (Ki) is variable. The aim of this study was to examine the impact of the fraction of victim drug metabolized by a particular CYP (fmCYP) and the inhibitor absorption rate constant (ka) on prediction accuracy.
Methods
Drug–drug interaction studies involving inhibition of CYP2C9, CYP2D6 and CYP3A4 (n = 115) were investigated. Data on fmCYP for the probe substrates of each enzyme and ka values for the inhibitors were incorporated into in vivo predictions, alone or in combination, using either the maximum hepatic input or the average systemic plasma concentration as a surrogate for [I]. The success of prediction (AUC ratio predicted within twofold of in vivo value) was compared using nominal values of fmCYP = 1 and ka = 0.1 min−1.
Results
The incorporation of fmCYP values into in vivo predictions using the hepatic input plasma concentration resulted in 84% of studies within twofold of in vivo value. The effect of ka values alone significantly reduced the number of over-predictions for CYP2D6 and CYP3A4; however, less precision was observed compared with the fmCYP. The incorporation of both fmCYP and ka values resulted in 81% of studies within twofold of in vivo value.
Conclusions
The incorporation of substrate and inhibitor-related information, namely fmCYP and ka, markedly improved prediction of 115 interaction studies with CYP2C9, CYP2D6 and CYP3A4 in comparison with [I]/Ki ratio alone.
Introduction
Drug–drug interactions resulting from inhibition of CYP-mediated metabolism can lead to serious toxicities, and have resulted in a number of compounds being withdrawn from the market. In recent years there has been an increased use of various in vitro systems used to detect CYP inhibition, which is qualitatively a useful tool. However, the extrapolation of these in vitro data to ultimately provide a quantitative in vivo prediction is problematic, and at present there is no comprehensive strategy that allows for the identification of particular drugs at risk from an inhibitory interaction [1–5].
In human in vivo interaction studies, the degree of interaction is expressed as the ratio of the area under the plasma concentration–time curve (AUC) in the presence and absence of an inhibitor. For many, but not all, cases this involves multiple oral dosing and the assumption is made that a new steady state is achieved. Also, for simplicity other conditions are commonly assumed: the victim drug is administered orally, cleared exclusively by the liver by way of a single metabolic pathway and the ‘well-stirred’ liver model applies. The AUC ratio is related to the ratio of the metabolic intrinsic clearance (CLint) as described by equation 1. The drug concentration in vivo is usually much lower than the Km value and the mechanism of inhibition (competitive or noncompetitive) is not relevant; therefore, equation 1 is valid for both inhibition types.
(1)where [I] is the inhibitor concentration available to the enzyme and subscript i indicates the presence of the inhibitor.
We have previously constructed a database of 146 studies to evaluate the prediction of drug–drug interactions involving reversible inhibition of CYP2C9, CYP2D6 and CYP3A4 [6]. In this analysis, we evaluated the utility of the [I]/Ki ratio by using various inhibitor plasma concentrations as surrogates for [I]. Results from this database analysis showed that the greatest change in AUC was observed for CYP3A4 (approximately 24-fold increase), followed by CYP2D6 (approximately 11-fold increase), with a fivefold AUC increase for CYP2C9 studies. The interaction studies involved nine different substrates for CYP2C9, 13 substrates for CYP2D6, with 18 substrates for CYP3A4 as shown in Figure 1, together with the predicted relationship based on equation 1. Using the maximum hepatic input concentration as [I] together with the in vitro Ki value was found to be the most successful method for categorizing CYP inhibitors and for identifying true negative drug–drug interactions. Although false negatives were eliminated, several false positives were evident and most true positives were markedly over-predicted (Figure 1). It was concluded that this generic approach provides only an initial discriminating screen, since there are a number of specific factors related to both the substrate and inhibitor that will affect the in vivo predictions.
Relationship between the observed AUC ratio and the [I]in/Ki ratio for 146 drug–drug interactions involving CYP2C9 (○), CYP2D6 (▵) and CYP3A4 (•). The line shown is the theoretical relationship based on equation 1. The shaded areas represent the regions corresponding to negative and positive drug–drug interactions as defined by the borderlines of an AUC ratio of 2 and an [I]/Ki of 1 [2]
Predictions made using equation 1 assume that the fraction of substrate metabolized by way of the inhibited CYP pathway ( fmCYP) is equal to 1. However, parallel pathways of metabolism and renal clearance of unchanged drug will affect the fmCYP and, consequently, the predicted degree of interaction, as even minor changes in the fmCYP value (e.g. from 1 to 0.98) may alter predictions significantly [7]. Equation 2 can be used in the prediction of the in vivo AUC ratio when fmCYP values are known and the other CYP pathways involved in the metabolism of the substrate are not subject to inhibition [7, 8]. Previously, we have demonstrated a substantial improvement in the quantitative predictions of drug–drug interactions involving CYP2D6 substrates using equation 2 rather than equation 1 [7].
(2)Previously [6], we investigated the use of the maximum hepatic inhibitor concentration at the inlet to the liver ([I]in). Calculation of this parameter (equation 3) relies on information on hepatic blood flow (QH), inhibitor dose (D), fraction absorbed from the gastrointestinal tract ( fa), the absorption rate constant (ka) to provide an absorption term and the average systemic plasma concentration ([I]av).
(3)In vivo clinical studies frequently do not report ka values; in the absence of this information and in order to avoid false-negative prediction and obtain the largest [I]in, it has been suggested that maximum ka of 0.1 min−1 is appropriate, assuming the gastric emptying is the rate limiting step for absorption [9].
The aim of the present study was to extend the previous database analysis [6] on 146 reversible drug–drug interaction studies and investigate the impact of substrate- and inhibitor-related parameters, namely fmCYP and ka, on the prediction accuracy, using either [I]av or [I]in together with published Ki data. Values for fmCYP were assigned for the commonly used substrate probes for CYP2C9, CYP2D6 and CYP3A4. In addition, ka values were estimated for each CYP inhibitor and the significance of these values on the [I]in value and in vivo predictions were assessed. The effects of fmCYP and ka, alone and in combination, have been analysed in order to maximize the drug–drug interaction prediction accuracy.
Methods
Data collection
Drug–drug interaction studies involving the reversible inhibition of CYP2C9, CYP2D6 and CYP3A4 (n = 146) were obtained from published literature [6]. The degree of interaction in each study was expressed as the fold increase in the AUC in the presence of an inhibitor, compared with the control study. In vitro Ki values for the CYP inhibitors involved in the above studies were also collected from the literature. In most cases in vitro data were available for the same substrate as used in the in vivo study, and when several human liver microsomal studies had been conducted, average Ki values were used for the prediction. If there were no available in vitro data involving the in vivo substrate in question, then in vitro data from alternative, well-established probe substrates of that particular enzyme were used [6]. For example, in the absence of in vitro studies involving fluconazole and phenytoin, the Ki value obtained with (S)-warfarin was used.
Values of fmCYP for each substrate were assigned using various literature data for a subset of 115 studies from the original database. The fmCYP value for the CYP2C9 substrate tolbutamide was obtained by calculating the difference between the urinary recovery of metabolites in both the presence and absence of the CYP2C9 selective inhibitor sulphaphenazole (phenocopying). Phenotyping data obtained from extensive and poor metabolizers of CYP2D6 were used to calculate the fmCYP values for these substrates [7]. A similar rationale was used to calculate the CYP2C19 contribution to phenytoin clearance and hence the fmCYP value for CYP2C9. For warfarin, an fmCYP value was calculated from a combination of urinary recovery of metabolites, biliary excretion and the recovery of unchanged drug as previously documented [10].
The fmCYP values are shown in Table 1; as the assignment of fmCYP values for CYP3A4 substrates was problematic, a range is shown for certain substrates. For all nine CYP3A4 substrates, the fraction excreted unchanged in urine is available and this provided an initial value for fmCYP based on the assumption that all metabolism is mediated via CYP3A4. In some cases this may be an upper estimate and further clarification is required. For the three benzodiazepines (midazolam, triazolam or alprazolam) this was achieved by adopting a regression approach [7] using equation 2 and the AUC ratio and [I]/Ki ratio for each substrate dataset (n = 8–16) to obtain an average fmCYP. Figure 2 shows the example of midazolam. The regression approach was also used for nifedipine and quinidine (n = 5 and 6). For the other CYP3A4 substrates (felodipine, nisoldipine, simvastatin and lovastatin) the number of studies available was more limited and fmCYP values were obtained by ranking the AUC ratio (using either data from itraconazole or ketoconazole studies) relative to midazolam and applying this factor to the midazolam fmCYP. For the predictions of the AUC ratio the lower values of fmCYP for the CYP3A4 substrates were used.
| CYP | Substrate | f mCYP * | References |
|---|---|---|---|
| 2C9 | Tolbutamide | 0.80 | [16] |
| S-warfarin | 0.87 | [10] | |
| Phenytoin | 0.75 | [17, 18] | |
| 2D6 | Desipramine | 0.88 | See [7] |
| Propafenone | 0.76 | ||
| Tolterodine | 0.94 | ||
| Encainide | 0.86 | ||
| Metoprolol | 0.83 | ||
| Mexiletine | 0.49 | ||
| Imipramine | 0.46 | ||
| Propranolol | 0.37 | ||
| 3A4 | Midazolam | 0.99, 0.94 | [24–26] |
| Triazolam | 0.98, 0.92 | [24] | |
| Alprazolam | 0.80 | [20] | |
| Nifedipine | 0.71 | [24, 27, 28] | |
| Nisoldipine | 0.99 | [27] | |
| Felodipine | 0.99, 0.81 | [24, 27, 29, 30] | |
| Quinidine | 0.76 | [21, 24, 27, 31] | |
| Simvastatin | 0.99 | [24] | |
| Lovastatin | 0.99 | [27, 32] |
- * When two values are shown, the higher value is derived from renal excretion data, whereas the lower value is obtained by regression/ranking and is that used in further predictions.
Determination of fmCYP for midazolam. Relationship between the AUC ratio observed in vivo and [I]in/Ki ratio for 10 drug–drug interactions involving midazolam as the victim drug
Data analysis
As described previously [6], the database analyses revealed that the inhibitor concentration was frequently not reported in an in vivo study, and when information was available in the same subjects, various concentrations were quoted (average, maximum or minimum). In order to standardize procedures, these concentrations were estimated from literature pharmacokinetic parameters. The average systemic plasma concentration after repeated oral administration ([I]av), and the maximum hepatic input concentration ([I]in) were calculated as in equations 4 and 3, respectively [9].
(4)In equation 4, F and τ represent the fraction of dose systemically available and dosing interval, respectively, of the inhibitor used in the in vivo interaction study. For the purpose of this analysis using equation 3, the fa value was taken as 1, assuming that the inhibitors were completely absorbed from the gastrointestinal tract, the ka value was initially assumed to be 0.1 min−1 (the maximum rate of gastric emptying) [9] and the blood-to-plasma concentration ratio and hepatic blood flow taken as unity and 1610 ml min−1, respectively.
The fmCYP data collected were incorporated into the prediction of AUC ratio using equation 2 for both [I]av and [I]in for all three CYPs and compared with initial in vivo predictions based on equation 1 for different [I]. In order to obtain more realistic ka estimates, values were calculated for each inhibitor using the time to reach maximum plasma concentration (Tmax) and the elimination rate constant (k) as shown in equation 5 (the latter values collected from published literature data). For a number of inhibitors (n = 5/10 for CYP2C9, n = 11/18 for CYP2D6 and n = 7/14 for CYP3A4), this pharmacokinetic information was unavailable; therefore a value of 0.01 min−1 was assigned. The calculated ka values for the inhibitors are listed in Table 2.
| Inhibitor | CYP enzyme | k a (min−1) | Reference |
|---|---|---|---|
| Sulphaphenazole | 2C9 | 0.030 | [9] |
| Fluconazole | 2C9, 3A4 | 0.061 | [33] |
| Ketoconazole | 2C9, 3A4 | 0.013 | [34] |
| Itraconazole | 3A4 | 0.020 | [35] |
| Quinidine | 2D6, 3A4 | 0.014 | [36] |
| Fluoxetine | 2D6, 3A4 | 0.009 | [37] |
| Fluvoxamine | 2C9, 2D6, 3A4 | 0.008 | [38] |
| Sertraline | 2C9, 2D6 | 0.007 | [37] |
| Citalopram | 2D6 | 0.024 | [37] |
| Nifedipine | 3A4 | 0.056 | [39] |
(5)Refined ka values (calculated from inhibitor pharmacokinetics or an assumed value of 0.01 min−1) were incorporated into in vivo predictions for all three CYP enzymes for [I]in, either alone or in combination with fmCYP information, using equations 1 and 2, respectively. The success of prediction (within twofold of in vivo value) was compared with the previous database analysis ( fmCYP = 1 and ka = 0.1 min−1). A twofold threshold value was selected on the basis of previous consensus reports [2, 11] for a significant increase in AUC ratio with a corresponding [I]/Ki ratio of unity.
The [I]/Ki ratio was calculated for each of the in vivo interaction studies using the various inhibitor concentrations described previously. Some inhibitors such as fluoxetine and itraconazole [6] have an active metabolite that also has inhibitory activity towards the same CYP enzyme. For these studies, the [I]/Ki ratio was calculated for the both the parent and the metabolite, the values were then added [12]. Out of the three itraconazole metabolites reported by Isoherranen et al.[13] (hydroxy-, keto- and N-desalkyl-itraconazole), only the contribution of hydroxy-itraconazole was included in the prediction, consistent with the previous database analysis [6].
The change in AUC ratio in vivo was plotted against the AUC ratio predicted using the various parameters and predictions within twofold of the in vivo AUC ratio were considered successful. The bias of drug–drug interaction prediction was assessed from the geometric mean of the ratio of predicted and actual value (average-fold error, afe). The mean squared prediction error (mse) (difference between the predicted and observed in vivo value) and the root mean squared prediction error (rmse) provided a measure of precision for the prediction of the drug–drug interaction studies using including [I]/Ki, appropriate ka and fmCYP values, both individually and in combination [14, 15].
(6)
(7)
(8)Results
From the original database [6] a subset of 115 studies was created for which fmCYP data on the in vivo probe substrate were available. The drug–drug interaction studies selected (n = 21 for CYP2C9, n = 40 for CYP2D6 and n = 54 for CYP3A4) involved 23 different substrates and 42 inhibitors. The range of fmCYP values was from 0.75 to 0.87 (CYP2C9), 0.37 to 0.94 (CYP2D6) and 0.71 to 0.99 (CYP3A4) (see Table 1). 3, 4 illustrate the effect of the fmCYP values on the prediction of AUC ratio for 115 drug–drug interaction studies (based on equation 2) using either [I]in or [I]av, respectively. The data in Figure 3A,B show that an improvement in the prediction accuracy is observed for each of the three CYP enzymes by incorporating the fmCYP values for in vivo predictions based on the [I]in. The number of studies within the twofold range of the in vivo value increased by 24, 38 and 28% for CYP2C9, CYP2D6 and CYP3A4, respectively (Table 3) and there was a corresponding reduction in the bias and increase in precision.
Relationship between the AUC ratio observed in vivo and the AUC ratio predicted for 115 drug–drug interaction studies involving CYP2C9 (○), CYP2D6 (▵) and CYP3A4 (•). The plots represent predictions using the maximum hepatic input concentration—equation 1 (A), incorporating both the fmCYP, equation 2 (B), refined ka value (C) and both fmCYP and ka(D). Solid line represents line of unity, whereas dashed lines represent the twofold limit in prediction accuracy. The shaded areas represent the regions corresponding to negative and positive drug–drug interactions as defined by the borderlines of an AUC ratio of 2 and an [I]/Ki of 1 [2]
Relationship between the observed AUC ratio in vivo and AUC ratio predicted for 115 drug–drug interaction studies involving CYP2C9 (○), CYP2D6 (▵) and CYP3A4 (•). The plots represent predictions using the average systemic total drug plasma concentration ([I]av) (A), and incorporating fmCYP data (B). The solid line represents line of unity, whereas dashed lines represent the twofold limit in prediction accuracy
| CYP | Prediction accuracy | [I]in* | [I]in with fmCYP | [I]in with refined ka | [I]in with fmCYP and refined ka |
|---|---|---|---|---|---|
| 2C9 | Over-predictions | 11 | 6 | 9 | 5 |
| Under-predictions | 0 | 0 | 1 | 1 | |
| Within twofold limit | 10 | 15 | 11 | 15 | |
| 2D6 | Over-predictions | 19 | 3 | 7 | 0 |
| Under-predictions | 0 | 1 | 2 | 5 | |
| Within twofold limit | 21 | 36 | 31 | 35 | |
| 3A4 | Over-predictions | 23 | 8 | 13 | 3 |
| Under-predictions | 0 | 0 | 5 | 8 | |
| Within twofold limit | 31 | 46 | 36 | 43 | |
| Total | % within twofold limit | 54 | 84 | 68 | 81 |
| afe | 2.11 | 1.21 | 1.37 | 0.84 | |
| rmse | 144.2 | 4.8 | 75.6 | 2.95 |
- * fmCYP = 1 and ka = 0.1 min−1.
Figure 4 indicates that the incorporation of fmCYP data into the in vivo predictions based on [I]av has a similar but less substantial effect. The greatest improvement occurred for CYP2C9 with a 24% increase in the number of studies within the twofold limit of the in vivo value. Incorporation of fmCYP data reduced several over-predictions for both CYP2D6 and CYP3A4. However, incorporation of fmCYP for the [I]av prediction did not significantly improve the under-predictions obtained for CYP3A4 interactions; 30% of studies involving this enzyme were still classed as false-negative interactions (see Figure 4B).
The [I]in value represents the combination of the circulating systemic plasma concentration and the additional concentration occurring during the absorption phase. Figure 5 illustrates the relationship between the [I]in and [I]av values for the 115 data studies where ka is assumed to be 0.1 min−1. The contribution of the absorption term ka · fa · D/Qh to the slope of this relationship can be illustrated by considering three particular inhibitors (Table 4). Lowering the ka value from the maximum value (ka = 0.1 min−1) to literature-reported values reduces the relative ratio between the absorption and systemic contribution by 10–13-fold for ketoconazole and itraconazole, but has a minimal effect for fluconazole. However, the dose absorbed is the main contributor to the high [I]in values and for [I]av this factor is less apparent due to the effect of volume of distribution.
Relationship between [I]in and [I]av for 115 drug–drug interaction studies involving CYP2C9 (○), CYP2D6 (▪) and CYP3A4 (▴)
| Inhibitor | [I]av(µm) | k a (min-1) | Absorption term (µm) | Absorption term/ systemic term |
|---|---|---|---|---|
| Ketoconazole | 0.44 | 0.013 | 3.0 | 6.8 |
| 0.1 | 23.4 | 53 | ||
| Itraconazole | 0.12 | 0.02 | 3.5 | 29 |
| 0.1 | 17.6 | 147 | ||
| Fluconazole | 23.2 | 0.061 | 24.6 | 1.1 |
| 0.1 | 40.55 | 1.7 |
Refinement of ka values from literature information was possible for 10 inhibitors to provide new parameter values for 86 studies (Table 2); a ninefold range of ka values was observed, ranging from 0.007 to 0.06 min−1 for sertraline and fluconazole, respectively. For the remaining studies involving inhibitors for which no absorption information was available, a ka value of 0.01 min−1 was assigned as a reasonable estimate. Figure 3C shows the effect of incorporating these refined ka values into in vivo predictions for 115 drug–drug interaction studies; fmCYP was assumed to be 1 for these predictions. From the results in Figure 3 and Table 3, it can be seen that in comparison with the original analysis [6] where an arbitrary ka of 0.1 min−1 was used, the ka improves the prediction accuracy for all three CYP enzymes. The greatest effect was noted for CYP2D6, where a 25% increase in the number of studies within twofold of the in vivo value is observed. In addition, the use of refined ka values significantly reduced the number of over-predictions in comparison with the higher ka value (2.2- and 2.7-fold for CYP3A4 and CYP2D6 drug–drug interaction studies, respectively).
The incorporation of both fmCYP and ka resulted in the most successful prediction for all three CYPs, with a total of 81% of studies within twofold of the in vivo value (Figure 4D and Table 3). For these predictions, there was the least bias and improvement in precision, as judged by the statistical parameters afe and rmse (see Table 3).
Discussion
In a previous drug–drug interaction database analysis [6] we have shown the utility of [I]in in qualitative zoning of inhibitors, allowing the true negatives to be identified and eliminating false negatives. However, several false positives resulted and on a quantitative level the large over-prediction of true-positive effects was of concern (see Figure 1). The fact that this simple generic approach ignores specific substrate- or inhibitor-related properties no doubt contributes to a number of over-predictions of true-positive interactions. Therefore, this study focused on demonstrating the significance of fmCYP for the victim drug (previously explored for CYP2D6 [7]) and ka for the inhibitor on the drug–drug interaction prediction accuracy for 115 studies. In order to assess the impact of these particular parameters on the predicted AUC ratio, previously collated data were used [6], including the literature reported Ki values.
The range of fmCYP values obtained for each CYP enzyme, 0.75–0.87 (CYP2C9), 0.37–0.94 (CYP2D6) and 0.71–0.99 (CYP3A4), illustrates that more than one enzyme/clearance mechanism contributes to the elimination of most of the victim drugs under consideration. The use of fmCYP data in the assessment of AUC ratio corrected several false-positive predictions, as well as reducing the extent of over-predictions of true positives (Table 3). The improvement is most notable for the predictions using [I]in, where the percentage of studies within the twofold limit of the in vivo AUC ratio increased from 54 to 84%.
Predictions for both CYP2C9 and CYP3A4 substrates are markedly improved to a comparable extent to that reported previously for CYP2D6 [7]. While fmCYP values for the CYP2C9 substrates are relatively high, they are sufficiently less than 1 to benefit substantially from adopting equation 2 rather than equation 1. Incorporation of the renal clearance contribution for quinidine and alprazolam reduced the overestimation of the interactions with these CYP3A substrates by three- and 30-fold, respectively, whereas the impact of fmCYP was of less significance for the other CYP3A4 substrates (fmCYP range from 0.9 to 0.99). These findings extend the analysis on CYP2D6 substrates previously presented [7] and illustrate the general applicability of fmCYP in progressing drug–drug interaction predictions to a valuable quantitative level.
A number of approaches were employed to obtain fmCYP in this study. For all the CYP2D6 substrates, comparison of phenotyping data in extensive and poor metabolizers of CYP2D6 was used [7]. As previously discussed [7], the phenotyping approach will provide the most unequivocal method for establishing the importance of a particular cytochrome P450 in the clearance of a drug. A good alternative for polymorphic enzymes is ‘phenocopying’, that is from the difference between the urinary recovery of metabolites in both the presence and absence of a selective inhibitor. We were able to use this approach for the CYP2C9 substrate tolbutamide using sulphaphenazole [16]. For another CYP2C9 substrate phenytoin, it is known that CYP2C19 also contributes to its clearance [17] and the availability of phenotyping data allowed calculation of the contribution of the latter CYP [18] and hence a fmCYP value for CYP2C9 and phenytoin. A fmCYP value for CYP2C9 and warfarin has been estimated by Kunze and Trager [10] using a combination of information on the urinary recovery of metabolites, biliary excretion and the recovery of unchanged drug. This level of detail is not commonly available, even for probe substrates.
For CYP3A4 substrates, estimation of fmCYP is problematic for several reasons, including the lack of selective inhibitors (ketoconazole and itraconazole being only selective at low concentrations) and complexities of multisite binding [19]. In this study, initial values were obtained from estimates of total metabolism calculated indirectly from urinary recovery of unchanged drug. These values are high (Table 1), which is consistent with the extensive use of several of these drugs (the benzodiazepines and calcium channel blockers) as selective probes. However, whether the metabolism of these substrates is completely mediated by CYP3A4 activity is debatable; thus these fmCYP values should be regarded as upper estimates. In the cases of alprazolam and quinidine, the importance of renal clearance is well established [20, 21], resulting in fmCYP values of ≤ 0.8. For the remaining seven substrates this method resulted in fmCYP values of ≥ 0.98. Simulations using equation 2 have shown AUC ratios to be very sensitive to small changes in fmCYP between values of 0.8 and 1 [7]; quinidine and alprazolam are the only CYP3A4 substrates outside this range. Most studies available used either midazolam, triazolam, alprazolam, quinidine or nifedipine, allowing a regression approach to be adopted based on equation 2 to obtain an average fmCYP for these five drugs (see Figure 2 for midazolam). For quinidine and alprazolam there was good agreement between the regression and corrected renal excretion values. For felodipine, nisoldipine, simvastatin and lovastatin, due to the limited number of studies available, fmCYP values were obtained by an alternative approach of ranking (either itraconazole or ketoconazole AUC ratios) relative to midazolam. Despite the limitations of these methods and the uncertainity of the absolute values of fmCYP obtained for the CYP3A4, 3, 4 indicate good predictions for these substrates, comparable to those for CYP2C9 and CYP2D6.
Extending this work to drug–drug interactions involving victim drugs that are not established CYP probes will rely on an estimate of fmCYP. Most drugs have several enzymes contributing to their elimination and the key information needed is the relative importance of particular enzymes to those drug pathways, i.e. fmCYP in contrast to the fraction metabolized by a particular pathway (often obtained via a radiolabel study). The importance of this type of specific information is being increasingly realized and various approaches have been recently summarized [11, 22]. The impact of hepatic transporters on drug clearance may also be an important consideration. However, the success apparent with CYP probe substrates described here, as well as theoretical relationships [7], would indicate that even approximate fmCYP values may markedly improve a prediction.
The use of [I]in relies on an input term for the hepatic portal vein plasma concentration calculated from equation 3. Predictions based on these [I] values, however, do result in a significant number of over-predictions or false-positive interactions [6]. One of the possible limitations of this approach is the use of the theoretical maximum value of 0.1 min−1 for the ka, which represents the maximum rate of gastric emptying [9]. Refinement of this parameter resulted in the ka values 2–14-fold lower than the initial estimates as shown in Table 2 for 10 CYP inhibitors investigated in the current study. Incorporation of refined ka reduced the relative contribution of the absorption term in comparison to the systemic term in the [I]in value up to 13-fold, as illustrated for itraconazole in Table 4. In addition, the ka value may vary with dose of inhibitor and the food intake (e.g. ketoconazole [23]), affecting the [I]in estimate and consequently the predicted AUC ratio.
Refined ka values reduce the number of over-predictions observed for all three CYPs (Table 3). Predictions using [I]in when either a realistic ka or fmCYP value were incorporated individually predicted 68–84% of the interactions within twofold of in vivo value (comparable lack of bias). However, incorporation of fmCYP improved the precision of the drug–drug interaction assessment (sevenfold lower rmse), substantially more than with the use of refined ka values (see Table 3).
The use of [I]in incorporating both fmCYP and refined ka values resulted in the most successful prediction overall (see Figure 3D). A total of 81% of studies were within the twofold limit of the in vivo value and this represents an increase of 30% in comparison with the qualitative zoning assumptions (ka = 0.1 min−1 and fmCYP = 1) previously described [6]. Minimal bias and high precision of the predictions were achieved (Table 3).
The accurate prediction of an in vivo drug–drug interaction is critically dependent on the inhibitor concentration used in equations 1 and 2. It is impossible to measure this concentration directly within the human liver and for this reason there are many conflicting reports about which inhibitor concentration to use in prediction, whether it is the systemic or portal vein concentration, total or unbound plasma concentration or the liver concentration. There have been many attempts to make an assessment of the concentration within the liver, with varying degrees of success [3, 9, 15]. Although in the present study the most successful predictions result from using a total drug concentration term ([I]in) with fmCYP and refined ka values, there are still a number of falsely predicted interactions. The possibility of an interaction in the gut wall may be significant for certain substrates and has not been included in this approach. Another factor that can influence the in vivo prediction is experimental variability in the generation of the in vitro data. The Ki values used in the current analysis are obtained from a variety of published literature sources and it would be valuable to explore whether standardization of the in vitro assessment would further improve prediction. This consideration is particularly pertinent for CYP3A4 Ki values, and a recent study has explored the importance of substrate selection and substitution for this enzyme [19].
In summary, we have demonstrated that incorporation of fmCYP values for the victim drug markedly improves prediction of 115 drug–drug interactions compared with the use of the [I]/Ki ratio alone. In addition to fmCYP, inclusion of realistic ka values to refine estimates of [I]in provides the most useful estimate of [I] and results in the most successful predictions as judged by a lack of bias and a high level of precision.
This work was funded by a consortium of pharmaceutical companies (AstraZeneca, Bristol Myers Squibb, GlaxoSmithKline, F. Hoffmann La Roche, Novartis, Pfizer and Servier) within the Centre for Applied Pharmacokinetic Research at the University of Manchester. H.S.B. was financially supported by a Bristol Myers Squibb studentship.
