Prediction of herb-drug interactions involving consumption of furanocoumarin-mixtures and cytochrome P450 1A2-mediated caffeine metabolism inhibition in humans

Abstract

Herb-drug interactions (HDI) has become important due to the increasing popularity of natural health product consumption worldwide. HDI is difficult to predict as botanical drugs usually contain complex phytochemical-mixtures, which interact with drug metabolism. Currently, there is no specific pharmacological tool to predict HDI since almost all in vitro-in vivo-extrapolation (IVIVE) Drug-Drug Interaction (DDI) models deal with one inhibitor-drug and one victim-drug. The objectives were to modify-two IVIVE models for the prediction of in vivo interaction between caffeine and furanocoumarin-containing herbs, and to confirm model predictions by comparing the DDI predictive results with actual human data. The models were modified to predict in vivo herb-caffeine interaction using the same set of inhibition constants but different integrated dose/concentration of furanocoumarin mixtures in the liver. Different hepatic inlet inhibitor concentration ([I]_H) surrogates were used for each furanocoumarin. In the first (hybrid) model, the [I]_H was predicted using the concentration-addition model for chemical-mixtures. In the second model, the [I]_H was calculated by adding individual furanocoumarins together. Once [I]_H values were determined, the models predicted an area-under-curve-ratio (AUCR) value of each interaction. The results indicate that both models were able to predict the experimental AUCR of herbal products reasonably well. The DDI model approaches described in this study may be applicable to health supplements and functional foods also.

1. Introduction

Regulatory bodies require the investigation of drug-drug interactions (DDI) prior to new-drug-candidate approval (Canada, 1998, EMEA, 2010, USFDA, 2012). Based on regulatory guidelines, new-drug-applicants are required to follow a stepwise protocol, which includes the investigation of major metabolizing enzymes inhibition, most notably cytochrome (CYP) P450 enzymes, in an effort to reduce unnecessary clinical trials and post-marketed drug-withdrawals. Several DDI prediction models have been developed by researchers and adopted by regulatory bodies (Mayhew et al., 2000, Wang et al., 2004, Einolf, 2007a, Einolf, 2007b, Fahmi et al., 2008, USFDA, 2012). Until recently, DDI prediction is based mainly on reversible CYP enzyme inhibition mechanisms and has been carried out routinely as part of drug preclinical studies (Fahmi et al., 2008). If irreversible inhibition were to occur, this would result in underestimation of the magnitude of DDI (Einolf, 2007a, Einolf, 2007b). Co-administration of a pharmaceutical drug and an herbal product, with bioactive constituents that interfere with drug metabolite(s) formation, might significantly alter the pharmacokinetics of the victim-drug. The outcome of such alterations may result in serious clinical consequence and/or deaths (Ebbesen et al., 2001).

Predictive DDI models have been developed with increased accuracy in prediction not only for reversible DDI but also for irreversible DDI. For example, Mayhew et al. (2000) reported a predictive DDI model that involves an irreversible inhibitor. Fahmi et al. (2008) proposed a combination model to sum up the inhibitory effects of reversible and irreversible enzyme inhibition along with enzyme induction. Despite these improvements, the DDI predictive model usually addresses interaction between one victim-drug and one inhibitor, which is not always the case in herb-drug interactions (HDI) as herbal products usually consist of a complex mixture of bioactive chemicals, which may interact with the victim-drug.

To implement a predictive DDI model involving time-dependent inhibitors (TDI), or the mechanism-based inhibitors (MBI), requires information on in vitro inactivation kinetics of the victim-drug as well as in vivo hepatic inlet inhibitor concentration ([I]_H). In the case of an herbal extract, the challenge is to estimate the [I]_H or an integrated dose of the bioactive chemicals of the herb in the liver. Wang et al. (2004) had modified the model of Mayhew et al. to account for a multitude of inhibitors, which probably can be applied to a mixture of bioactive inhibitors in an herbal extract. The DDI model of Wang et al. requires detailed pharmacokinetics information on each inhibitor in the herbal extract. These can be time-consuming and expensive at the early drug development stage. In this study, we propose to use the concentration-addition (CA) model (Safe, 1998, ATSDR, 2004, Alehaideb et al., 2019) to calculate an integrated dose/concentration for the furanocoumarin-mixture in the liver. The integrated dose is then used to calculate the area-under-curve-ratio (AUCR) with the Mayhew et al. model.

The “victim-drug” in this study is caffeine. A popular drug with adverse-health-effects upon abusive consumption (Dews, 1982). Caffeine is also an ideal probe to measure in vivo CYP1A2 activity (Doehmer et al., 1992, Miners and Birkett, 1996). The pharmacokinetics of caffeine in humans has been studied extensively (Kot and Daniel, 2008). The “perpetrators” in the present study are the linear furanocoumarins (Fig. 1) which are chemical isomers and congeners found in Apiaceae, Leguminosae, Moraceae, and Rutaceae plant families (Diawara and Trumble, 1997). Previous in vitro liver microsomal studies have shown that the main metabolic pathway of 8-methoxypsoralen (8-MOP) and 5- methoxypsoralen (5-MOP) is the oxidative ring-opening of the furan structure to form an epoxide, or an intermediate electrophilic reactive metabolite, which binds covalently to human CYP proteins (Fouin-Fortunet et al., 1986, Tinel et al., 1987, Mays et al., 1990, John et al., 1992). Other linear furanocoumarins such as isopimpinellin (ISOP) (Kang et al., 2011) and psoralen (Zhuang et al., 2013) also have been shown to be TDI or MBI in human liver microsomes (HLM) or recombinant human CYP1A2 expressed in yeast.

Fig. 1.

The chemical structures of linear furanocoumarins of psoralen and derivatives.

Our previous studies have shown that detectable levels of 8-MOP, 5-MOP, and ISOP in nine herbal medicines (Alehaideb et al., 2017). We also have shown four of the furanocoumarin-containing herbs significantly reduced the oral clearance of caffeine in human volunteers (Alehaideb et al.; Alehaideb and Matou-Nasri, 2021). Furthermore, the aforementioned furanocoumarins were found to be potent TDI or MBI of CYP1A2 isozyme (Alehaideb et al., 2021). The objectives of this study were: (1) to modify the predictive DDI models of Mayhew et al. and Wang et al. for in vivo herb-caffeine interaction using in vitro caffeine metabolism data in HLM and in vivo C_max of furanocoumarins in humans, and (2) to determine the accuracy of Mayhew et al. and Wang et al. DDI predictive models by comparing predicted caffeine AUCR with experimental AUCR from Alehaideb et al. (2021).

2. Methods

2.1. Source of plant products

The plant products used in this study were obtained from North American commercial suppliers as follows: Ammi majus L. seeds (A. majus) purchased from EverWilde (Fallbrook, CA), Angelica archangelica L. roots (A. archangelica), Apium graveolens seeds (A. graveolens S), Pimpinella anisum L. seeds (P. anisum), and Ruta graveolens L. leaves (R. graveolens) were purchased from Mountain Rose (Eugene, OR), Apium graveolens L. leaves (A. graveolens L) and Petroselinum crispum (Mill.) Fuss leaves (P. crispum) were purchased from A1SpiceWorld (Glen Head, NY). Angelica pubescens Maxim. roots (A. pubescens) purchased from Spring Wind (San Francisco, CA), Cnidium monnieri (L.) Cusson seeds (C. monnieri) were purchased from Health and Wellness House (Duncan, BC). The plant products were further authenticated chromatographically as detailed in our previous publication (Alehaideb et al., 2017).

2.2. Using DDI models to predict herb-caffeine interactions.

In the present study, caffeine is the victim-drug and the furanocoumarin bioactive in the herbs are the perpetrators or inhibitors as mentioned earlier. As both caffeine and furanocoumarin inhibitors are metabolized by the same CYP1A2 enzyme, metabolic inhibition may occur in humans after co-administration. Indeed, the furanocoumarins have been shown previously to be TDI or MBI of CYP1A2 (Alehaideb et al., 2021). We have used two different DDI models to predict the inhibition of caffeine metabolism by furanocoumarin-containing herbs namely the Mayhew et al. (hybrid) DDI model and Wang et al. DDI model.

2.2.1. Using Mayhew et al. (hybrid) DDI model to predict caffeine-herb interactions.

Mayhew et al. first proposed a DDI model involving the destruction of CYP enzyme by the reactive metabolite(s) of the inhibitor or the mechanism-based DDI model as follows:

Equation 1: The IVIVE DDI prediction model of Mayhew et al. (2000).

$$\mathrm{A}\mathrm{U}\mathrm{C}\mathrm{R}=\frac{{\mathrm{A}\mathrm{U}\mathrm{C}}_{\mathrm{I}}}{\mathrm{A}\mathrm{U}\mathrm{C}}=\frac{1}{1+\left(\frac{{\mathrm{f}}_{\mathrm{m}}}{\frac{{[\mathrm{I}]}_{\mathrm{H}}\times {\mathrm{k}}_{\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{c}\mathrm{t}}}{{\mathrm{k}}_{\mathrm{d}\mathrm{e}\mathrm{g}}\times \left({[\mathrm{I}]}_{\mathrm{H}}+{\mathrm{K}}_{\mathrm{I}}\right)}}\right)+\left(1-{\mathrm{f}}_{\mathrm{m}}\right)}$$ (1)

where AUC is the area-under-curve of caffeine with no herbal extract pre-treatment, AUC_I is the area-under-curve of caffeine with herbal extract pre-treatment, k_deg is the first-order in vivo degradation rate constant for CYP enzyme, K_I represents the equilibrium dissociation constant for the inactivator, k_inact is the maximum rate of enzyme inactivation at saturating concentrations of inhibitor, f_m is the fraction of metabolic pathway, and [I]_H represents the in vivo hypothetical intra-hepatic concentration(s) of the integrated dose/concentration of furanocoumarin inhibitors. Thus, the model of Mayhew et al. required the determination of K_I, k_inact, k_deg and [I]_H for the chemical marker (CM) furanocoumarin of 8-MOP only. It should be noted that the 8-MOP was chosen as the CM in the CA model as it displayed the lowest IC₅₀ value among the furanocoumarins in this study.

The theoretical background of mechanism-based model has been described in great detail elsewhere (Mayhew et al., 2000). The model is able to predict quantitatively the inhibition of drugs if the following conditions are met: (a) the well-stirred liver conditions are met, (b) the in vitro inhibition constants are applicable to in vivo interaction, and (c) the [I]_H value is lower than the K_I inhibition constant value.

An integrated dose/concentration of the furanocoumarin mixture in the liver, [I]_H was predicted using the CA model (Safe, 1998, ATSDR, 2004, Alehaideb et al., 2019) with 8-MOP as the CM as follows:

Equation 2: The concentration-addition (CA) model approach.

$$\mathrm{F}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{n}\mathrm{o}\mathrm{c}\mathrm{o}\mathrm{u}\mathrm{m}\mathrm{a}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{C}\mathrm{M}\mathrm{e}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}=\sum \mathrm{C}\mathrm{I}\times \mathrm{R}\mathrm{P}\mathrm{F}$$ (2)

where CI is the mass concentration of the individual furanocoumarin in the chemical mixture and RPF is the relative potency of each chemical in comparison to 8-MOP potency.

The CA model is applicable to the present study because: (a) the furanocoumarin inhibitors are chemical isomers and congeners that exhibit similar biological mechanisms, (b) the dose-inhibition curves of individual furanocoumarins are parallel to one another as the slopes of the dose-inhibition curves of 8-MOP, 5-MOP, and ISOP were 2.0, 1.6, and 1.9 respectively (Alehaideb et al., 2021), and (c) the inhibitory potencies of individual furanocoumarins are additive in nature (Alehaideb and Matuo-Nasri, 2021).

2.2.2. Using Wang et al. Model to predict caffeine-herb interaction.

Instead of calculating a single [I]_H with the CA model as seen in the hybrid Mayhew et al. model approach (Safe, 1998, ATSDR, 2004), individual furanocoumarin [I]_H were calculated separately before adding them together to yield the AUCR value. Thus, Wang et al. model required the determination of the K_I, k_inact, k_deg, and [I]_H values for each individual furanocoumarin inhibitor as follows:

Equation 3: The IVIVE DDI prediction model of Wang et al. (2004).

$$\mathrm{A}\mathrm{U}\mathrm{C}\mathrm{R}=\frac{{\mathrm{A}\mathrm{U}\mathrm{C}}_{\mathrm{I}}}{\mathrm{A}\mathrm{U}\mathrm{C}}=\frac{1}{1+\left(\sum \frac{{\mathrm{f}}_{\mathrm{m}}}{\frac{{[\mathrm{I}]}_{\mathrm{H}}\times {\mathrm{k}}_{\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{c}\mathrm{t}}}{{\mathrm{k}}_{\mathrm{d}\mathrm{e}\mathrm{g}}\times {\mathrm{K}}_{\mathrm{I}}}}\right)+\left(1-{\mathrm{f}}_{\mathrm{m}}\right)}$$ (3)

Since caffeine undergoes negligible first-pass metabolism (Kalow and Tang, 1993), the DDI model of Wang et al. was modified for the present study by excluding the intestinal metabolism term from the original equation.

Fig. 2 summarizes the experimental procedure and validation steps of the herb-caffeine interaction studies. Model-predicted AUCR was compared with experimental AUCR in previous pharmacokinetic studies (Alehaideb et al., 2021) in order to validate the IVIVE models for herb-caffeine interaction prediction.

Fig. 2.

Flow-chart summary of experimental procedures and validation steps in the use of drug-drug interaction models to predict herb-drug interaction.

2.2.3. IVIVE DDI model input parameters.

The predictive models of Mayhew et al. and Wang et al. used a combination of in vitro caffeine metabolism inhibition parameters (Alehaideb et al., 2021) and in vivo furanocoumarin [I]_H concentrations to predict quantitative impairment of in vivo caffeine clearance due to co-administration of caffeine and furanocoumarin-containing herbs. As shown in Eqs. (1), (2), in vivo caffeine clearance impairment does not depend on the concentration of caffeine but is dependent on the concentrations of furanocoumarin inhibitors at the site of metabolism, the liver. As the intra-hepatic concentrations of furanocoumarin inhibitors could not be determined by direct experiment, four different furanocoumarin dose surrogates (i.e., maximum total plasma concentration (C_max,PT), maximum unbound plasma concentration (C_max,PU), maximum total liver concentration level (C_max,LT), and maximum unbound liver concentration (C_max,LU) were used to calculate the [I]_H for the predictive models.

2.2.3.1. In vitro caffeine metabolism inhibition data.

The in vitro experimental inhibition parameters we obtained using radiolabeled caffeine and pooled HLM as described in detail in our previous study (Alehaideb et al., 2021). Briefly, the IC₅₀ values were obtained using serial dilutions of pure furanocoumarin incubated with radiolabeled caffeine, HLM, and NADPH cofactor. The radiolabeled metabolites were collected by solid-phase extraction and counted by scintillation. The inactivation constants were experimentally measured using the two-step dilution with pre-incubation time points ranging from 0.5 to 4 min. The IC₅₀ of 8-MOP, 5-MOP, and ISOP were 0.09, 0.13, and 0.29 µM, respectively. The corresponding K_I values were 0.78, 3.73, and 4.48 µM, respectively. The k_inact values were 0.17, 0.35, and 0.65 min⁻¹, respectively (Alehaideb et al., 2021). The k_deg value of CYP1A2, 0.0003 min⁻¹, was obtained from Faber and Fuhr (2004). The f_m value for caffeine in humans, 0.95, was taken from Kalow and Tang (1993). Thus, only the C_max-derived [I]_H remained to be determined in the IVIVE models.

2.2.3.2. Furanocoumarin C_max extrapolation based on dose-C_max relationship

Table 1 lists the body weights (BW) of the volunteers and the doses of individual furanocoumarins administered as calculated from our previous publications (Alehaideb et al., 2017, Alehaideb et al., 2021). These furanocoumarin doses were used as the administered doses in predicting the C_max,PT of 5-MOP and 8-MOP in the plasma of humans based on allometric extrapolation (Gabrielsson and Weiner, 2006). The C_max,PT of ISOP was the average of the 5-MOP and 8-MOP values. The following is the detailed description of predicting C_max,PT, C_max,PU, C_max,LT, and C_max,LU concentration levels in humans:

• (a)The C_max,PT of 8-MOP and 5-MOP were obtained by simple and direct allometric extrapolation from the experimental data of Schäfer-Korting and Korting, 1982, Stolk et al., 1981 studies, respectively, using the method described in Gabrielsson and Weiner (2006). Briefly, the C_max values were for each volunteer and furanocoumarin was directly extrapolated using the Ln-Ln plot of furanocoumarin doses (mg/kg BW) and C_max values (ng/L) as seen in Fig. 3. Fig. 3a shows the administered dose versus serum C_max,PT plot of 8-MOP, the range is 482.0 to 822.0 µg/kg BW as it appears in the graph 3a. (Schfifer-Korting and Korting, 1982). Fig. 3b shows the administered dose versus serum C_max,PT plot of 5-MOP, the range is 600.0 to 1200.0 µg/kg BW as it appears in the graph 3b (Stolk et al. 1981). Despite our best effort, we were unable to find any animal or human data for proper allometric extrapolation of ISOP. Instead, we assumed the C_max,PT values of ISOP to be the average of mean 8-MOP and mean 5-MOP extrapolated values.
• (b)C_max,PU was derived from C_max,PT by multiplying the latter with the unbound fractions (fup) of 8-MOP or 5-MOP in the plasma; they were 17.0 (±7.4) and 3.6 (±2.2) percent respectively (Veronese et al., 1978, Artuc et al., 1979, Pibouin et al., 1987, Makki et al., 1991, Muret et al., 1993a). No information was found for ISOP; therefore, an average of 8-MOP and 5-MOP values was used.
• (c)The C_max,LT was derived by multiplying C_max,PT with the liver:plasma partition coefficient (Pt:p) which was calculated as follows:

Table 1.

A summary of body weight and individual furanocoumarin doses for human volunteers in the caffeine pharmacokinetic studies.

Plant Name and Part	na	Human Furanocoumarin Oral Dose
		BW b	8-MOP b	5-MOP b	ISOP b
		kg	µg/kg BW	µg/kg BW	µg/kg BW
A. majus seeds	4	77.5 ± 17.1	260.9 ± 73.4	58.2 ± 16.4	611.9 ± 172.1
A. archangelica roots	5	73.2 ± 17.6	42.2 ± 11.2	25.5 ± 6.8	39.3 ± 10.4
C. monnieri fruits	5	75.8 ± 7.4	28.2 ± 3.0	71.3 ± 7.5	18.6 ± 2.0
R. graveolens leaves	4	72.0 ± 6.9	56.3 ± 5.7	22.4 ± 2.3	12.3 ± 1.3
A. pubescens roots	4	75.0 ± 8.2	4.1 ± 0.5	5.2 ± 0.6	n.d. c
A. graveolens seeds	4	79.3 ± 5.8	2.7 ± 0.2	2.1 ± 0.1	29.9 ± 2.1
A. graveolens Leaves	2	81.5 ± 3.5	1.5 ± 0.1	29.9 ± 1.3	1.2 ± 0.1
P. crispum leaves	2	81.5 ± 3.5	n.d. c	4.2 ± 0.2	n.d.c
P. anisum seeds	4	75.8 ± 8.5	2.1 ± 0.3	n.d. c	n.d. c

^an = number of human volunteers.

^bExpressed as mean ± SD.

^cn.d. = not detected in herbal extract.

Fig. 3.

Ln-Ln plot of (a) 8-MOP dose versus C_max,PT data from Schfifer-Korting and Korting (1982) and (b) Ln-Ln plot of 5-MOP dose versus C_max,PT data from Stolk et al. (1981).

Equation 4: Non-adipose tissue:plasma partition model of Poulin and Theil (2002).

$$P_{t:p}=\frac{P_{o/W}\times (V_{\mathit{nt}}+0.3\times V_{\mathit{pht}})+(V_{\mathit{wt}}+0.7\times V_{\mathit{pht}})}{P_{o/W}\times (V_{\mathit{np}}+0.3\times V_{\mathit{php}})+(V_{\mathit{wp}}+0.7\times V_{\mathit{php}})}\times \frac{f_{\mathit{up}}}{f_{\mathit{ut}}}$$ (4)

where P_o/w is the n-octanol/water partition for non-ionized inhibitor, V_nt is the fraction weight of neutral lipids in liver tissue, V_pht is the fraction weight of phospholipids in liver tissue, V_wt is the fraction weight of water in liver tissue, V_np is the fraction weight of neutral lipids in plasma, V_wp is the fraction weight of water in plasma, V_php is the fraction weight of phospholipids in plasma, f_up is the fraction unbound in plasma, and f_ut is the fraction unbound in liver tissue.

The P_o/w of 8-MOP, 5-MOP, and ISOP were 120.2, 128.8, and 169.8 respectively; they were obtained from the Advanced Chemistry Development I-lab databases (ACD, 2015). The values of V_nt, V_pht, V_wt, V_np, V_wp, and V_php were 0.035, 0.025, 0.751, 0.004, 0.945, and 0.002 respectively; they were also obtained from Poulin and Theil (2000).

• (d)The C_max,LU was derived by multiplying the C_max,LT by the calculated unbound fraction (f_ut) in tissue. The f_ut was determined using the following equation:

Equation 5: Unbound tissue fraction model of Poulin and Theil (2000).

$${\mathrm{f}}_{\mathrm{u}\mathrm{t}}=\frac{1}{1+((\frac{1-{\mathrm{f}}_{\mathrm{u}\mathrm{p}}}{{\mathrm{f}}_{\mathrm{u}\mathrm{p}}})\times 0.5)}$$ (5)

2.3. Human caffeine pharmacokinetic studies

The in vivo pharmacokinetic data for caffeine metabolism inhibition due to consumption of furanocoumarin-containing herbal products were reported in our previous publications in great detail (Sheriffdeen et al., 2019, Alehaideb et al., 2021, Alehaideb and Matou-Nasri, 2021). Briefly, eligible adult male volunteers, aged between 20 and 40 years, provided written consent and documentation is kept in a safe location prior experiments. The volunteers were dosed with 200 mg caffeine twice: with and without prior treatment with an aqueous extract of one selected herbal medicine. Saliva samples were collected at time-points ranging between 0 and 48 h. Caffeine and internal standard (benzotriazole) were separated and measured chromatographically using an isocratic method with an ultra-violet detector as described in Perera et al. (2010) with slight modifications. The salivary caffeine concentrations were converted into plasma concentrations using a conversion factor of 0.79 ± 0.2 which is based on a study involving 12 healthy adult volunteers which were given 250–300 mg caffeine dose (Zylber-Katz et al., 1984, Fuhr et al., 1993). The human plasma caffeine AUC from zero to infinity (AUC_0-inf) for each volunteer was measured twice and the experimental AUCR was calculated using the PKsolver software (Zhang et al., 2010). These studies were approved by Simon Fraser University (Office of Research Ethics) with approval number 2012 s0565 and registration number ISRCTN83028296.

2.4. Data and statistical analysis.

Data plotting and extrapolation were performed using GraphPad Prism Software version 5.04 (San Diego, CA). Statistical analysis was performed using Microsoft Excel software. Model-predicted AUCR were reported as mean ± standard deviation (SD). Herb-caffeine interaction occurred when the mean AUCR was ≥2.0. No interaction occurred when the mean AUCR was less than 2.0 (Einolf, 2007a, Einolf, 2007b). The geometric mean fold-error (GMFE) (Eq. (6)) was also used to assess the accuracy of model prediction by equal weighting under-predictions and over-predictions. The model that predicted perfectly would give a GMFE value of 1; GMFE value between 1 and 2-fold is considered to be accurate.

Equation 6: The geometric mean fold-error (GMFE) (Obach et al., 1997).

$$\mathrm{G}\mathrm{M}\mathrm{F}\mathrm{E}={10}^{\frac{\sum \left|\mathrm{l}\mathrm{o}\mathrm{g}\frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d}\mathrm{H}\mathrm{D}\mathrm{I}}{\mathrm{A}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{a}\mathrm{l}\mathrm{H}\mathrm{D}\mathrm{I}}\right|}{\mathrm{n}}}$$ (6)

where n is the number of predictions for each herb.

3. Results and discussion

Table 2, Table 3, Table 4, Table 5 display the experimental and model-predicted AUCR of caffeine in humans as a result of pre-treatment with an herbal extract. The variability of the experimental AUCR is large but consistent with the large variation of 8-MOP (Herfst and De Wolff, 1982) and 5-MOP (Ehrsson et al., 1994) concentrations observed in the serum. Relatively high experimental AUCR were observed in volunteers pre-treated by A. majus, A. archangelica, A. pubescens, C. monnieri, or R. graveolens as seen in Table 2, Table 3, Table 4, Table 5. These results are consistent with the relatively high levels of furanocoumarins found in the herbal extracts as seen in Table 1. Moreover, the presence of osthole, a CYP1A2 inhibitor (Yang et al., 2012), in A. pubescens may contribute to the high AUCR of this herbal extract. Similarly, the relatively high AUCR of C. monnieri might be due to osthole in addition to high levels of 8-MOP, 5-MOP, and ISOP. Despite only low levels of furanocoumarins have been found in P. crispum; mean AUCR in the volunteers is noticeable. This might be due to the presence of flavones such as apigenin in the extract (Meyer et al., 2006), which is a CYP1A2 inhibitor (Peterson et al., 2006).

Table 2.

A comparison of model-predicted C_max,PT-based AUCR and experimental AUCR in humans due to herbal extract pretreatment.

Botanical Name	Actual AUCR			Predicted AUCR
				Wang et al. (2004)					Mayhew et al. (2000)
	Mean	±	SD	Mean	±	SD	Predictive a	GMFE b	Mean	±	SD	Predictive a	GMFE b
A.majus seeds	4.7	±	##	19.9	±	##	+	4.3	17.9	±	##	+	3.9
A.archangelica roots	2.3	±	##	6.3	±	##	+	2.7	2.0	±	##	+	1.2
C.monnieri fruits	2.6	±	##	11.9	±	##	+	5.0	1.8	±	##	─	1.5
R.graveolens leaves	1.7	±	##	2.7	±	##	─	1.6	2.0	±	##	─	1.4
A.pubescens roots	1.9	±	##	1.0	±	##	+	1.7	1.0	±	##	+	1.8
A.graveolens seeds	1.3	±	##	3.4	±	##	─	2.8	1.1	±	##	+	1.4
A.graveolens leaves	1.2	±	##	2.8	±	##	─	2.4	1.2	±	##	+	1.2
P.crispum leaves	1.6	±	##	1.0	±	##	+	1.6	1.0	±	##	+	1.6
P.aniseum seeds	1.1	±	##	1.0	±	##	+	1.2	1.0	±	##	+	1.2
							Average	2.6				Average	1.7

^aBased on the twofold rule; (+) predictive and (─) not predictive.

^bGeometric meanfold error; (≤2) indicate accurate prediction and (>2) is less accurate.

Table 3.

A comparison of model-predicted C_max,PU-based AUCR and experimental AUCR in humans due to herbal extract pretreatment.

Botanical Name	Actual AUCR			Predicted AUCR
				Wang et al. (2004)					Mayhew et al. (2000)
	Mean	±	SD	Mean	±	SD	Predictive a	GMFE b	Mean	±	SD	Predictive a	GMFE b
A.majus seeds	4.7	±	##	19.3	±	##	+	4.1	11.1	±	##	+	2.3
A.archangelica roots	2.3	±	##	1.7	±	##	─	1.4	1.1	±	##	─	2.0
C.monnieri fruits	2.6	±	##	3.3	±	##	+	1.6	1.1	±	##	─	2.2
R.graveolens leaves	1.7	±	##	1.2	±	##	+	1.4	1.2	±	##	+	1.4
A.pubescens roots	1.9	±	##	1.0	±	##	+	1.8	1.0	±	##	+	1.8
A.graveolens seeds	1.3	±	##	1.3	±	##	+	1.4	1.0	±	##	+	1.4
A.graveolens leaves	1.2	±	##	1.1	±	##	+	1.2	1.0	±	##	+	1.2
P.crispum leaves	1.6	±	##	1.1	±	##	+	1.5	1.0	±	##	+	1.6
P.aniseum seeds	1.1	±	##	1.0	±	##	+	1.2	1.0	±	##	+	1.2
							Average	1.7				Average	1.7

^aBased on the twofold rule; (+) predictive and (─) not predictive.

^bGeometric meanfold error; (≤2) indicate accurate prediction and (>2) is less accurate.

Table 4.

A comparison of model-predicted C_max,LT-based AUCR and experimental AUCR in humans due to herbal extract pretreatment.

Botanical Name	Actual AUCR			Predicted AUCR
				Wang et al. (2004)					Mayhew et al. (2000)
	Mean	±	SD	Mean	±	SD	Predictive a	GMFE b	Mean	±	SD	Predictive a	GMFE b
A.majus seeds	4.7	±	##	20.0	±	##	+	4.3	19.1	±	##	+	4.1
A.archangelica roots	2.3	±	##	10.2	±	##	+	4.4	2.7	±	##	+	1.4
C.monnieri fruits	2.6	±	##	15.8	±	##	+	6.7	2.8	±	##	+	1.5
R.graveolens leaves	1.7	±	##	4.5	±	##	─	2.7	3.1	±	##	─	1.9
A.pubescens roots	1.9	±	##	1.1	±	##	+	1.6	1.0	±	##	+	1.8
A.graveolens seeds	1.3	±	##	6.3	±	##	─	5.2	1.2	±	##	+	1.4
A.graveolens leaves	1.2	±	##	4.6	±	##	─	3.8	1.5	±	##	+	1.2
P.crispum leaves	1.6	±	##	1.1	±	##	+	1.6	1.0	±	##	+	1.6
P.aniseum seeds	1.1	±	##	1.0	±	##	+	1.2	1.0	±	##	+	1.2
							Average	3.5				Average	1.8

^aBased on the twofold rule; (+) predictive and (─) not predictive.

^bGeometric meanfold error; (≤2) indicate accurate prediction and (>2) is less accurate.

Table 5.

A comparison of model-predicted C_max,LU-based AUCR and experimental AUCR in humans due to herbal extract pretreatment.

Botanical Name	Actual AUCR			Predicted AUCR
				Wang et al. (2004)					Mayhew et al. (2000)
	Mean	±	SD	Mean	±	SD	Predictive a	GMFE b	Mean	±	SD	Predictive a	GMFE b
A.majus seeds	4.7	±	##	19.8	±	##	+	4.3	16.6	±	##	+	3.6
A.archangelica roots	2.3	±	##	3.9	±	##	+	1.6	1.4	±	##	─	1.6
C.monnieri fruits	2.6	±	##	8.6	±	##	+	3.6	1.4	±	##	─	1.7
R.graveolens leaves	1.7	±	##	1.7	±	##	+	1.3	1.6	±	##	+	1.3
A.pubescens roots	1.9	±	##	1.0	±	##	+	1.8	1.0	±	##	+	1.8
A.graveolens seeds	1.3	±	##	2.3	±	##	─	1.9	1.0	±	##	+	1.4
A.graveolens leaves	1.2	±	##	1.3	±	##	+	1.2	1.0	±	##	+	1.2
P.crispum leaves	1.6	±	##	1.0	±	##	+	1.6	1.0	±	##	+	1.6
P.aniseum seeds	1.1	±	##	1.0	±	##	+	1.2	1.0	±	##	+	1.2
							Average	2.0				Average	1.7

^aBased on the twofold rule; (+) predictive and (─) not predictive.

^bGeometric meanfold error; (≤2) indicate accurate prediction and (>2) is less accurate.

Wang et al. IVIVE model predicted a ≥ 2-fold increase in mean caffeine AUCR after pre-treating the volunteers with A. majus seeds or C. monnieri fruits extract regardless of C_max used to derive the [I]_H value. However, the interaction due to A. archangelica roots pre-treatment was not predicted successfully using C_max,PU values as seen in Table 3. The AUCR derived from C_max,PU and C_max,LU appeared to have fewer false positive results in the remaining 6 herbs on the list as seen in Table 2, Table 3, Table 4, Table 5. The average GMFE for C_max,PU and C_max,LU were 1.7 and 2.0, as seen in Table 3, Table 5 respectively, indicating C_max,PU and C_max,LU yield more accurate AUCR results than C_max,PT and C_max,LT which had average GMFE value of 2.6 and 3.5, respectively, as seen in Table 2, Table 4. These results are consistent with the larger number of false positive results in C_max,PT and C_max,LT as seen in Table 2, Table 4. The actual AUCR versus predicted AUCR correlation plot confirms C_max,PU and C_max,LU are better dose surrogates to predict caffeine AUCR than C_max,PT and C_max,LT as seen in Fig. 4a.

Fig. 4.

Correlation of (a) Wang et al. (2004) model-predicted AUCR and (b) of Mayhew et al. (2000) model-predicted AUCR against actual AUCR using different [I]_H surrogate values.

In contrast, when the IVIVE model of Mayhew et al. was used to predict the AUCR of caffeine in humans pre-treated by the herbal extracts, only C_max,LT predicted the experimental AUCR closely as seen in Table 4. An exception was R. graveolens leaves which overpredicted the experimental AUCR. The average GMFE were similar for C_max,LU C_max,PT and C_max,PU at 1.7 or 1.8 as seen in Table 2, Table 3, Table 4, Table 5. However, C_max,PU and C_max,LU generally underpredicted the experimental AUCR with the exception of A. majus seeds which was always over-predicted regardless of the DDI model used. Thus, C_max,LT yielded the most accurate prediction for caffeine metabolism inhibition with average GMFE of 1.8 as seen in Table 4. The actual AUCR versus predicted AUCR correlation plot also confirms C_max,LT is a better dose surrogate to predict caffeine AUCR than C_max,LU, C_max,PT and C_max,PU as seen in Fig. 4b.

Despite no agreement is reached on the most accurate [I]_H surrogate for DDI predictions, previous studies have reported accurate DDI prediction by using either total C_max or unbound C_max in the liver and plasma to estimate [I]_H (Grimm et al., 2009). For example, Ito et al. (2004; 2005) have reported that C_max,LT yields the most accurate prediction for reversible-based inhibition DDI. This is consistent with the suggestion of Brown et al. (2006) that incorporating protein binding into the predictive model did not improve reversible-based DDI prediction. In contrast, Obach et al., 2006, Fahmi et al., 2009 have concluded that C_max,LU is most accurate in reversible-based DDI prediction. Blanchard et al. (2004) have reported C_max,PU provides the most accurate DDI predictions based on reversible inhibition. For irreversible-based inhibition predictions, Obach et al., 2007, Fahmi et al., 2009 have suggested C_max,PU is the most accurate dose surrogate for DDI prediction. In contrast, both Ito et al., 2003, Shardlow et al., 2011 reported C_max,LU provided the most accurate DDI predictions. Nevertheless, the studies do agree that systemic and liver C_max values, both total and free (unbound), should be considered.

Although both Wang et al. and Mayhew et al. models were able to predict the empirical AUCR of caffeine, Wang et al. model is more predictive than Mayhew et al. model as the latter is implemented with a single, composite [I]_H value which is less accurate than a combination of individual [I]_H values in Wang et al. model. The predictability of the IVIVE model is also affected by the kinetic parameters as follows: (a) the k_deg value of hCYP1A2. The 0.00030 min⁻¹ k_deg used in the present study is based on the 38 h t_1/2 of a tobacco smoking cessation study (Faber and Fuhr, 2004). Mayhew et al. has reported a k_deg value of 0.00083 min⁻¹ which is derived from rats. If this k_deg was used in the present study, both predictive models would underestimate the AUCR of caffeine, (b) in vitro kinetic inhibition parameters such as IC₅₀, KI, and k_inact were derived using pooled HLM from multiple donors. This also appears to improve the accuracy of our predictions, and (c) caffeine as the victim drug has simplified model prediction by eliminating the need to account for parallel metabolic pathways by other CYP isoforms and urinary excretion of unchanged caffeine. As a result, the uncertainty involved in the AUCR calculation is greatly reduced.

The following are few of the main limitations, assumptions, and observations of the present study as follows: (a) the data used to establish the 8-MOP dose–response curves for allometric extrapolation are based on psoriasis patients and these raised concern that plasma 8-MOP levels in healthy and psoriasis subjects might be different. However, Shephard et al. (1999) has shown that similar free 8-MOP fractions are found in the plasma of psoriasis and healthy volunteers. Also, Muret et al. (1993b) have shown that the fractions of free 5-MOP in the serum of healthy and psoriasis volunteers do not differ significantly. Together, these results suggest that systemic 8-MOP and 5-MOP levels do not differ significantly in healthy and psoriasis subjects, (b) in the present study, the pharmacokinetic values of individual furanocoumarins in an herbal extract and pure furanocoumarins are assumed to be similar. This assumption might be incorrect since pharmacokinetic interaction between individual furanocoumarins in an herbal extract may occur, (c) previous studies have shown that both 8-MOP (Mays et al., 1987, Apseloff et al., 1990) and ISOP (Baumgart et al., 2005) induce CYP1A2 enzyme activity in rats after multiple dosing. In contrast, Tantcheva-Poór et al. (2001) dosed humans with 0.6 mg of 8-MOP/kg/day and found no significant induction in caffeine clearance even after a week of treatment. In the present study, the inductive effect of 8-MOP and ISOP is ignored as only a single dose of herbal extract is administered 3 hrs before the caffeine pharmacokinetic studies, and it is unlikely CYP1A2 can be induced with a short pretreatment period, and (d) both modified models tend to over-predict at a certain [I]_H value reaching the maximum prediction AUCR value equal twenty. However, this observation has minimal effect as the objective of these models is to measure the AUCR in sufficient accuracy but more importantly is to detect DDI with minimal false-negative predictions.

4. Conclusions

Using IVIVE model methods to predict HDI is an ongoing research program in our laboratories. Our goal is to develop a reliable and simple prediction tool for HDI. Both modified DDI models of Wang et al. and Mayhew et al. predicted reasonably well with the latter hybrid model requiring fewer experiments utilizing the concentration-addition approach which could save time and cost. The study also suggests linear furanocoumarins, such as 8-MOP, 5-MOP and ISOP, are responsible for the inhibition of caffeine metabolism in humans after consuming herbal extracts containing furanocoumarin derivatives. Finally, the described modeling approaches in this study may also be applicable to other natural products, health supplements, and functional foods.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.