Predictability of human exposure by human‐CYP3A4‐transgenic mouse models: A meta‐analysis

Abstract

First‐in‐human dose predictions are primarily based on no‐observed‐adverse‐effect levels in animal studies. Predictions from these animal models are only as effective as their ability to predict human results. To narrow the gap between human and animals, researchers have, among other things, focused on the replacement of animal cytochrome P450 (CYP) enzymes with their human counterparts (called humanization), especially in mice. Whereas research in humanized mice is extensive, the emphasis has been particularly on qualitative rather than quantitative predictions. Because the CYP3A4 enzyme is most involved in the metabolism of clinically used drugs, most benefit was expected from CYP3A4 models. There are several applications of these mouse models regarding in vivo CYP3A4 functionality, one of which might be their capacity to help improve first‐in‐human (FIH) dose predictions for CYP3A4‐metabolized drugs. To evaluate whether human‐CYP3A4‐transgenic mouse models are better predictors of human exposure compared to the wild‐type mouse model, we performed a meta‐analysis comparing both mouse models in their ability to accurately predict human exposure of small‐molecule drugs metabolized by CYP3A4. Results showed that, in general, the human‐CYP3A4‐transgenic mouse model had similar accuracy in the prediction of human exposure compared to the wild‐type mouse model, suggesting that there is limited added value in humanization of the mouse Cyp3a enzymes if the primary aim is to acquire more accurate FIH dose predictions. Despite the results of this meta‐analysis, corrections for interspecies differences through extension of human‐CYP3A4‐transgenic mouse models with pharmacokinetic modeling approaches seems a promising contribution to more accurate quantitative predictions of human pharmacokinetics.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

Whereas research in humanized mice is extensive, the emphasis has been particularly on qualitative rather than quantitative predictions.

• WHAT QUESTION DID THIS STUDY ADDRESS?

Are human‐CYP3A4‐transgenic mouse models better predictors of human exposure compared to the wild‐type mouse model for small‐molecule drugs metabolized by CYP3A4?

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

In general, the human‐CYP3A4‐transgenic mouse model had similar accuracy in the prediction of human exposure compared to the wild‐type mouse model, suggesting that there is limited added value in humanization of the mouse Cyp3a enzymes if the primary aim is to acquire more accurate first‐in‐human (FIH) dose predictions.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

Humanization of CYP3A enzymes alone is not enough to account for the misspecifications in prediction of human exposure in the context of FIH dosing. Interspecies differences consist of an interplay of many different processes that are vastly more complex. Modern data analysis approaches might help to exploit benefits of human‐CYP3A4‐transgenic mouse models.

INTRODUCTION

To ensure the safety of participants of first‐in‐human (FIH) clinical trials of new molecular entities, regulatory guidelines by both the US Food and Drug Administration and European Medicines Agency describe how to derive maximum recommended starting doses. ¹ , ² , ³ The aim is to predict a starting dose close to the therapeutic range, especially for anticancer drugs, in order to acquire phase I objectives (e.g., assessment of the pharmacodynamic and pharmacokinetic [PK] profile and drug tolerability) within a reasonable time, limiting the number of participants treated at subtherapeutic doses while minimizing toxicity at the initial dose. Predictions of the human PKs and pharmacodynamics of a drug are made based on in vitro assays and in vivo animal models prior to exposing humans. The recommended process involves determining the no‐observed‐adverse‐effect levels (NOAELs) in different animal species and converting the NOAEL of the most sensitive species to the human equivalent dose using allometric scaling. ¹ For anticancer drugs, the severely toxic dose in 10% of animals is commonly used. ⁴ Nevertheless, to predict human exposure and toxicity, animal models are only as effective as their ability to predict human results. Hence, the World Health Organization recommends a factor 10 safety margin over the NOAELs to allow for interspecies differences. This factor of 10 is constituted of the subfactors 2.5 and 4.0 for toxicodynamics and toxicokinetics, respectively. ⁵

In the late 1980s, researchers acquired the skills to genetically modify animal models by knocking out certain animal genes and replacing them with their human counterparts to better predict the human PKs, a process called humanization. In the field of PKs, the often‐observed inconsistency of metabolizing enzymes between species is a common target for humanization, especially in mice. ⁶ One main purpose for these models was to recognize risks and opportunities for in vivo human drug–drug and drug‐food interactions in in vivo mouse settings in a qualitative way. For instance, the human cytochrome P450 (CYP)3A enzymes deviate from the Cyp3as of mice. Mice express eight full‐length mouse Cyp3as and humans four CYP3As (CYP3A4, ‐5, ‐7, and ‐43). Despite differences, human CYP3A and mouse Cyp3a have broadly overlapping substrate specificity and tissue expression. Therefore, the biological function of all wild‐type mouse Cyp3as combined likely corresponds to the combined function of all human CYP3As. ⁷ However, because of intrinsic biological differences between these species (e.g., preferred diet), these functionalities are not necessarily identical. Therefore, the wild‐type model may not be the most appropriate model to investigate the PKs of drugs for which clearance is highly dependent on CYP3A4‐mediated metabolism. For instance, reliably studying drug–drug interactions would be practically impossible because compounds responsible for human CYP3A4 inhibition are not necessarily inhibitors for the mouse Cyp3as (and vice versa).

Because CYP3A4 is the enzyme most frequently involved in metabolism of many clinically used drugs, and often affected by inter‐ and intra‐individual differences in expression and activity, ⁸ multiple research groups developed humanized CYP3A4 mouse models to investigate the metabolism of CYP3A4. Although human‐CYP3A4‐transgenic mouse models have mainly been studied for the qualitative translational assessment of effects of the CYP3A4 enzyme in humans, few studies have focused on the quantitative predictability of the human‐CYP3A4‐transgenic mouse models over that of the wild‐type mouse model. Better quantitative predictions of human exposure could potentially help inform the FIH dose that still requires high margins of safety due to inaccuracy of animal models. ⁵ We here performed a meta‐analysis of the literature in which human‐CYP3A4‐transgenic mouse models were used to assess the PKs of small molecule drugs. Our aim with this meta‐analysis is to evaluate whether the human‐CYP3A4‐transgenic mouse models provide a better prediction of human exposure than the wild‐type mouse models.

METHODS

A literature search for publications presenting quantitative PK information after administration of a small‐molecule drug in a human‐CYP3A4‐transgenic animal model was last performed on the February 6, 2023, using PubMed. The meta‐analysis was performed according to the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses guidelines, except that screening and reporting was performed by only one reviewer. ⁹ The full PubMed search and inclusion and exclusion criteria are presented in Figure 1. PubMed was searched using the search term ((cytochrome P450 AND 3a OR 3a4) OR cyp3a4) AND (transgenic OR humanized) AND (pharmacokinetics OR exposure OR AUC OR C_max OR (peak concentration)). Exclusion criteria were the absence of useful plasma PK data, CYP3A5/7, chimeric liver transplant mice, reviews, absence of human PKs for drug, last measurable concentration (C _last) higher than 1/10th of the maximum plasma concentration (C _max), and no possibility for extrapolation until infinity and no or unknown dose proportionality. Extraction of the area under the concentration‐time curve extrapolated to infinity (AUC_0–inf) from the publication was performed. Either the C _last was at least 1/10th of the C_max or at least two observed concentrations after the C _max were available to calculate the elimination constant (k _e) and extrapolate the AUC_0–inf by dividing the C _last by the k _e (with the assumption that the studied drug had a single terminal elimination rate constant). Data from concentration‐time curves were extracted by means of PlotDigitizer ¹⁰ and AUC_0–inf calculated using the linear trapezoidal rule with extrapolation. Subsequently, the dose administered in mice was allometrically scaled to a human equivalent dose:

$$\text{Dose scaling factor}={\frac{{\mathrm{BW}}_{\text{human}}}{{\mathrm{BW}}_{\text{mouse}}}}^{\text{exponent}}$$ (1)

Where BW_human and BW_mouse represent the bodyweight of an average human and mouse, which were assumed to be 70 and 0.03 kg, respectively. The exponent represents the allometric scaling exponent of either 0.67 or 0.75. An exponent of 0.67 yields a function between BW and clearance that is similar to a linear function between body surface area and clearance. The exponent of 0.75 is used to describe interspecies differences in basal metabolic rate. ¹¹ , ¹² Because they are generally both used, we evaluated both. The AUC_0–inf corresponding to the human equivalent dose was extracted from literature. If no data regarding the AUC_0–inf were available for the human equivalent dose, the closest available dose with information regarding AUC_0–inf was extracted (the fold difference between the scaled and closest available dose ranged from 0.09 to 108; Table S1). The extracted AUC_0–inf and dose in this case were linearly scaled to the human equivalent dose under the assumption that the PKs of the drug of interest were dose proportional. Last, all AUC_0–inf units were converted to ng/mL h and mouse AUC_0–inf were compared to human AUC_0–inf for the accuracy in the prediction of human exposure. A schematic overview of the methods is presented in Figure 2.

FIGURE 1.

Flow diagram of literature search and article selection. *For 20 of the 53 AUC values, the C _last are unknown (all AUC_0–24h), however, after 24 h, most drugs in mice reach a concentration smaller than 1/10th of the C _max and therefore the study was assumed to fulfill the criteria and the AUC values were not excluded (no wild‐type mice PK was evaluated in these experiments). AUC, area under the plasma concentration‐time curve; C _last, last observed concentrations; C _max, peak concentrations; PK, pharmacokinetic.

FIGURE 2.

Schematic presentation of the methods. AUC_0–inf, area under the plasma concentration‐time curve until infinity; PK, pharmacokinetic; WT, wild‐type.

Comparison of the deviations of mouse AUC_0–inf from human AUC_0–inf for multiple drugs results in higher absolute errors for drugs that have higher AUC_0–inf despite a low relative error, which is more informative here. To give equal weights to the predictability of the mouse model for the human exposure for each compound we calculated the fold differences from the human AUC_0–inf. In order to calculate mean errors of the fold differences, normalization across fold differences smaller and larger than one are required. Fold differences were normalized using Equation 2:

$$\text{Normalized fold difference}={10}^{\left|{\log }_{10}\text{fold difference}\right|},$$ (2)

Subsequently, normalized fold differences were used to calculate the mean absolute error (MAE) and the root mean squared error (RMSE).

$$\mathrm{MAE}=\frac{1}{n}\sum _{i=1}^n\left|y_i-{\widehat{y}}_i\right|,$$ (3)

$$\text{RMSE}=\sqrt{\sum _{i=1}^n\frac{{\left({\widehat{y}}_i-y_i\right)}^2}{n}},$$ (4)

Where y _i represents the observation and ŷ _i the prediction.

The mouse model resulting in lower median errors and the least dispersion was considered a better predictor of the human exposure. Processing of the data and graphical and statistical diagnostics were performed with R (version 4.2.1).

RESULTS

The literature search resulted in 161 publications. Of these, 28 met our inclusion criteria and were used for the analysis. A flow diagram of the inclusion and exclusion is presented in Figure 1. After exclusion, only studies in mice were described in the remaining publications. We identified eight publications that described the development of human‐CYP3A4‐transgenic mouse models (original transgenic mouse models; Table 1). Two publications (Abe et al. and Cheung et al. ¹³ , ¹⁴ ) were not used to evaluate the PKs of any drug (also not in other publications) and three publications (Hasegawa et al., Kazuki et al. and Ma et al. ¹⁵ , ¹⁶ , ¹⁷ ) describing the original transgenic mouse models did not present drug PKs itself or did not meet the inclusion criteria (other publications did evaluate the PKs of drugs using these models). Furthermore, two crossbred models have been developed by Uehara et al. and Scheer et al. by crossbreeding previously developed human‐CYP3A4‐transgenic mouse models. ¹⁸ , ¹⁹ Within the 28 included publications, two publications (Damoiseaux et al. and Zhang et al. ²⁰ , ²¹ ) used modeling approaches to extrapolate human‐CYP3A4‐transgenic mouse model PKs to human PKs and were included for discussion. Fifty‐three AUC_0–inf were derived from the other 26 publications (Tables 2 and S1) containing 26 unique drugs administered in human‐CYP3A4‐transgenic mouse models developed by eight different mouse model developers (of which two crossbred). From 17 of the 26 publications, also the AUC_0–inf in wild‐type mice could be derived (for 19 of the 53 AUC_0–inf in human‐CYP3A4‐transgenic mice). Additionally, Table 3 presents human drug exposures reported in the literature after administration of a single dose of the drugs that were also evaluated in the mouse models, as well as to what extent they are metabolized by CYP3A4.

TABLE 1.

Publications describing human‐CYP3A4‐transgenic mouse models.

Model	Human P450 transgene structure	Promoter	Mouse gene knockout for humanization	References
CYP3A4‐transgenic	Gene (BAC)	Authentic	None	Granvil et al. 31
CYP3A4‐transgenic	cDNA	Human ApoE promoter	None	van Herwaarden et al. 32
CYP3A4‐humanized	cDNA	Human ApoE promoter or mouse villin promoter	Cyp3a‐null	van Herwaarden et al. 7
CYP3A4/3A7‐transgenic	Gene (BAC)	Authentic	None	Cheung et al. 14
CYP3A4/3A7‐transgenic	Gene (BAC)	Authentic (PXR humanized)	None	Ma et al. 17
CYP3A4/3A7‐humanized	Gene (modified BAC)	Authentic (PXR and CAR humanized)	Cyp3a‐null (except for Cyp3a13)	Hasegawa et al. 15
CYP3A4/3A5*3/3A7/3A43‐humanized	Gene (HAC)	Authentic (CYP3A5 not expressed)	Cyp3a‐null	Kazuki et al. 16
CYP3A4/3A5*1/3A7/3A43‐humanized	Gene (modified HAC)	Authentic (CYP3A5 expressed)	Cyp3a‐null	Abe et al. 13

Models developed by crossbreeding of human‐CYP3A4‐transgenic mouse models were not included. Information originates from Table 1, Bissig et al. ⁶

Abbreviations: ApoE, Apolipoprotein E; BAC, bacterial artificial chromosome; CAR, constitutive androstane receptor; HAC, human artificial chromosome; PXR, Pregnane X receptor.

TABLE 2.

Summary of all included publications presenting quantitative pharmacokinetic information after administration of a small‐molecule drug in a human‐CYP3A4‐transgenic mouse model.

References	Category a	Evaluated drugs	Mice model developer	Crossbred mice model developer	Wild‐type mice PK available?	Administered Dose (mg/kg) (oral unless indicated otherwise)	Reported AUC for transgenic mice	Reported AUC for transgenic mice wild‐type	AUC time interval	C last lower than 1/10th of C max?	AUC0–inf for transgenic mice (ng/mL h)	AUC0–inf for wild‐type mice (ng/mL h)
Choo, E. F., et al. 33	Application	Cobimetinib	Hasegawa et al.; Van Herwaarden et al. 2007	NA	Yes	5	0.701 ± 0.087 μM h (van Herwaarden); 3.95 ± 0.58 μM h (Hasegawa)	1.38 ± 0.25 μM h	0–24 h	Yes	372.4 (van Herwaarden); 2098.6 (Hasegawa)	733.2
Damoiseaux, D., et al. 34	Application	Lorlatinib	Van Herwaarden et al. 2007	NA	Yes	10	7.8 ± 1.3 μg/mL h; 9.2 ± 1.5 μg/mL h	13.3 ± 2.2 μg/mL h; 17.2 ± 3.0 μg/mL h	0–8 h; 0–inf	Yes	9200.0	17200.0
Granvil, C. P., et al. 31	Original	Midazolam	Granvil et al.	NA	Yes	2.5 (oral); 0.25 (i.v.)	8330 nmol/L min (oral); 8390 nmol/L min (i.v.)	13,800 nmol/L min (oral); 6530 nmol/L min (i.v.)	0–inf	Yes	45.0 (oral) 45.3 (i.v.)	74.5 (oral) 35.3 (i.v.)
Hasegawa, M., et al. 35	Application	Triazolam	Hasegawa et al.	NA	No	5	1210 ± 110 ng/mL h	NA	0–inf	Yes	1210.0	NA
Henderson, C. J., et al. 36	Application	Caffeine, debrisoquine, midazolam, tolbutamide, dabrafenib, sulfaphenazole, S‐Acenocoumarol, Hyperforin	Hasegawa et al.	Scheer et al.	No	10 dabrafenib; 3 midazolam; 25 osimertinib	21,925 ± 1687 ng/mL h dabrafenib; 462 ± 70 ng/mL h midazolam; 3753 ± 614 ng/mL h osimertinib (no PK for other drugs)	NA	0–25 h (dabrafenib); 0–inf (midazolam); 0–32 h (osimertinib)	Yes	21925.0 (dabrafenib) 462.0 (midazolam) 3753.0 (osimertinib)	NA
Kim, S., et al. 37	Application	Triazolam	Ma et al.	NA	No	4	1583 nM h	NA	0–6 h	No (1/8th C max)	569.7	NA
Kobayashi, K., et al. 38	Application	Triazolam	Kazuki et al.	NA	No	1	1050 ± 242 nM h; 448 ± 157 nM h	NA	0–inf	Yes	153.7	NA
Li, W., et al. 39	Application	Lorlatinib	Van Herwaarden et al. 2007	NA	Yes	10	11,701 ± 1274 ng/mL h	7585 ± 533 ng/mL h	0–8 h	No (half C max)	12557.3	10553.2
Li, W., et al. 40	Application	Lorlatinib	Van Herwaarden et al. 2007	NA	Yes	10	5792 ± 871 ng/mL h	10,542 ± 1067 ng/mL h	0–8 h	No (half C max)	7972.0	16699.1
Li, W., et al. 41	Application	Fisogatinib	Van Herwaarden et al. 2007	NA	Yes	10	4284 ± 600 ng/mL h	5072 ± 823 ng/mL h	0–4 h	No (1/5th C max)	5956.8	4753.4
Li, W., et al. 42	Application	Galunisertib	Van Herwaarden et al. 2007	NA	Yes	20	4095 ± 1262 ng/mL h	2706 ± 640 ng/mL h	0–1 h	Yes	4095.0	2706.0
Ly, J. Q., et al. 43	Application	Alprazolam, bosutinib, crizotinib, dasatinib, gefitinib, ibrutinib, regorafenib, sorafenib, triazolam, vandetanib	Hasegawa et al.; Van Herwaarden et al. 2007	NA	No	5	8.07 ± 3.47 Alprazolam, 0.41 ± 0.21 bosutinib, 0.616 ± 0.128 crizotinib, 0.105 ± 0.035 dasatinib, 0.89 ± 0.06 gefitinib, 0.174 ± 0.058 ibrutinib, 4.9 ± 1.53 regorafenib, 8.44 ± 3.22 sorafenib, 0.22 ± 0.046 triazolam, 10.2 ± 1.92 vandetinib (μM h, van Herwaarden) ‐‐‐‐‐‐‐‐‐‐ 3.1 ± 0.23 Alprazolam, 1.84 ± 0.35 bosutinib, 0.213 ± 0.08 crizotinib, 0.119 ± 0.08 dasatinib, 1.68 ± 0.39 gefitinib, 0.567 ± 0.105 ibrutinib, 2.67 ± 0.4 regorafenib, 8.02 ± 1.15 sorafenib, 1.15 ± 0.21 triazolam, 6.88 ± 1.55 vandetinib (μM h, Hasegawa)	NA	0–24 h	Unknown	2492.0 alprazolam, 217.5 bosutinib, 279.2 crizotinib, 53.7 dasatinib, 397.7 gefitinib, 74.9 ibrutinib, 2365.7 regorafenib, 3922.9 sorafenib, 75.5 triazolam, 4889.9 vandetinib (van Herwaarden) ‐‐‐‐‐‐‐‐‐‐ 957.3 alprazolam, 975.9 bosutinib, 94.6 crizotinib, 58.6 dasatinib, 750.8 gefitinib, 251.1 ibrutinib, 1289.1 regorafenib, 3727.7 sorafenib, 394.7 triazolam, 3298.3 vandetinib (Hasegawa)	NA
MacLeod, A. K., et al. 44	Application	Vemurafenib	Hasegawa et al.	Scheer et al.	No	50 and 100	294 μg/mL h (50 mg/kg), huCYP3A4/3A7; 432 μg/mL h (50 mg/kg), 559 μg/mL h (100 mg/kg), huPXR/huCAR/huCYP3A4/3A7	NA	0–inf	Yes	294,000 (50 mg/kg) huCYP3A4/3A7; 432,000 (50 mg/kg), 559,000 (100 mg/kg), huPXR/huCAR/huCYP3A4/3A7	NA
MacLeod, A. K., et al. 45	Application	Osimertinib	Hasegawa et al.	Scheer et al.	Yes	25	1144 ± 363 ng/mL h	998 ± 419 ng/mL h	0–24 h	Yes	1144.0	998.0
Martínez‐Chávez, A., et al. 46	Application	Ribociclib	Van Herwaarden et al. 2007	NA	Yes	20	1834 ± 490 ng/mL h	5901 ± 1760 ng/mL h	0–8 h	No (1/3rd C max)	2173.1	7443.9
Martínez‐Chávez, A., et al. 47	Application	Abemaciclib	Van Herwaarden et al. 2007	NA	Yes	10	3200 ± 791 nM h	7808 ± 1837 nM h	0–24 h	Yes	1622.4	3958.7
Miura, T., et al. 48	Application	S‐warfarin, Diclofenac	Hasegawa et al.	NA	No	0.5 warfarin; 10 diclofenac (both i.v.)	8.3 ± 2.4 nmol/mL h (S‐warfarin); 41.1 ± 7.5 nmol/mL h (diclofenac)	NA	0–inf	Yes	12165.6 (diclofenac) (S‐warfarin excluded)	NA
ML, F. M., et al. 49	Application	Niraparib	Van Herwaarden et al. 2007	NA	Yes	50	56,463 ± 10,785 ng/mL h	25,919 ± 6309 ng/mL h	0–24 h	Yes	18068.2	8294.1
Scheer, N., et al. 19	Crossbred	Midazolam	Hasegawa et al.	Scheer et al.	No	5	125 μg/mL min	NA	0–24 h	Yes	2083.3	NA
Uehara, S., et al. 18	Application	caffeine, warfarin, omeprazole, metoprolol, midazolam	Hasegawa et al.	Uehara et al.	No	10	13 ± 3 caffeine; 58 ± 11 warfarin; 0.022 ± 0.010 omeprazole; 0.038 ± 0.002 metoprolol; 0.098 ± 0.017 midazolam (μg/mL h)	NA	0–inf	Yes	20.0 (omeprazole) 100.0 (midazolam) (caffeine, warfarin and metoprolol excluded)	NA
van Herwaarden, A. E., et al. 32	Original	Midazolam, Cyclosporin A	Van Herwaarden et al. 2005	NA	Yes	30 midazolam; 20 cyclosporin A (both i.v.)	5.45 μg/mL h (midazolam); 24.3 μg/mL h (cyclosporin A)	11.7 μg/mL h (midazolam); 35.8 μg/mL h (cyclosporin A)	0–3 h (midazolam); 0–8 h (cyclosporin A)	Yes (midazolam); No 1/7th C max (cyclosporin A)	5450.0 (midazolam) 31800.0 (cyclosporine A)	11700.0 (midazolam) 48020.0 (cyclosporine A)
van Herwaarden, A. E., et al. 7	Original	Docetaxel	Van Herwaarden et al. 2007	NA	Yes	10 (i.v.)	976.9 ng/mL h	777 ng/mL h	0–8 h	Yes	976.9	777.0
van Hoppe, S., et al. 22	Application	Ibrutinib	Van Herwaarden et al. 2007	NA	Yes	10	832 ± 521 ng/mL h	431 ± 96.6 ng/mL h	0–8 h	Yes	832.0	431.0
van Waterschoot, R. A., et al. 50	Application	Triazolam	Van Herwaarden et al. 2007	NA	Yes	0.5	130 ± 19 μg/L h	194 ± 23 μg/L h	0–5.3 h	No (1/3rd C max)	176.2	NA
Wang, J., et al. 51	Application	Tivozanib	Van Herwaarden et al. 2007	NA	Yes	1	6227 ± 936 ng/mL h	4557 ± 683 ng/mL h	0–24 h	Yes	6227.0	4557.0
Yamazaki, H., et al. 52	Application	Midazolam	Hasegawa et al.	NA	Yes	10 (i.v.)	759 ± 431 μM min	536 ± 46 μM min	0–inf	Yes	4121.1	2910.3

Abbreviations: AUC_0–inf, from zero to infinite hour; AUC, area under the plasma concentration‐time curve; C _last, last observed concentrations; C _max, peak concentrations; NA, not applicable; PK, pharmacokinetics.

^a The column category consists of: original publications that describe the development of a human‐CYP3A4‐transgenic mouse model; crossbred, publications that describe the crossbreeding of a human‐CYP3A4‐transgenic mouse model; application, publications that describe pharmacokinetic experiments using a human‐CYP3A4‐transgenic mouse model.

TABLE 3.

Reported human drug exposure after administration of a single dose.

Drug (oral unless indicated otherwise)	CYP3A4 metabolized drug?	Degree of CYP3A4 mediated metabolism of drug	Drug class	Reference human PK data	Reported AUC for human	AUC time interval	C last lower than 1/10th of C max?	Calculated AUC0–inf
Abemaciclib	Yes	Extensively	L01EF03 (Antineoplastic agents)	Patnaik A, et al. (2016) 53	1270 ng/mL h (50 mg); 1880 ng/mL h (100 mg); 4010 ng/mL h (150 mg); 5220 ng/mL h (200 mg)	0–inf	Yes	NA
Alprazolam	Yes	Primarily	N05BA12 (Psycholeptics)	Friedman H, et al. (1991) 54	305 ng/mL h (1 mg)	0–50 h	No, 1/9th C max	316.6 ng/mL h
Bosutinib	Yes	Primarily	L01EA04 (Antineoplastic agents)	Abbas R, et al. (2011) 55	323 ng/mL h (100 mg)	0–inf	Yes	NA
Cobimetinib	Yes	Primarily	L01EE02 (Antineoplastic agents)	Rosen LS, et al. (2016) 56	1556 ng/mL h (40 mg) a ; 3112 ng/mL h (60 mg) a	0–24 h a	Yes	NA
Crizotinib	Yes	Primarily	L01ED01 (Antineoplastic agents)	Xu H, et al. (2015) 57	1260 ng/mL h (150 mg); 2192 ng/mL h (250 mg)	0–inf	Yes	NA
Cyclosporin A (i.v.)	Yes	Extensively	L04AD01 (Immunosuppressants)	Gupta SK, et al. (1990) 58	8799 ng/mL h (4 mg/kg, 64 kg)	0–24 h	Yes	NA
Dabrafenib	Yes	Primarily	L01EC02 (Antineoplastic agents)	Ouellet D, et al. (2013) 59	9858 ng/mL h (150 mg)	0–inf	Yes	NA
Dasatinib	Yes	Primarily	L01EA02 (Antineoplastic agents)	Christopher LJ, et al. (2008) 60	1151 ng/mL h (180 mg)	0–24 h	Yes	NA
Diclofenac (i.v.)	No	NA	M01AB55 (Anti‐inflammatory and antirheumatic products)	Leuratti C, et al. (2019) 61	5384 ± 1020 ng/mL h (75 mg; i.v. bolus)	0–inf	Yes	NA
Docetaxel (i.v.)	Yes	Primarily	L01CD02 (Antineoplastic agents)	Baker SD, et al. (2006) 62	3.41 μg/mL h (75 mg/m2; 1 h infusion)	0–inf	Yes	NA
Fisogatinib	Yes	Unknown	Unknown	Kim RD, et al. (2019) 63	24,420 ng/mL h (140 mg) a ; 128,564 ng/mL h (600 mg) a	0–24 h a	Yes	NA
Galunisertib	Unlikely	Unknown	Unknown	Ding X, et al. (2015) 64	3670 μg/L h (150 mg, solution)	0–inf	Yes	NA
Gefitinib	Yes	Partly	L01XX31 (Antineoplastic agents)	Ranson M, et al. (2002) 65	1147 ng/mL h (50 mg)	0–140 h	Yes	NA
Ibrutinib	Yes	Primarily	L01EL01 (Antineoplastic agents)	Tapaninen T, et al. (2020) 66	76.5 ng/mL h (140 mg)	0–inf	Yes	NA
Lorlatinib	Yes	Primarily	L01ED05 (Antineoplastic agents)	Patel M, et al. (2020) 67	7338 ng/mL h (100 mg)	0–inf	Yes	NA
Midazolam	Yes	Extensively	N05CD08 (Psycholeptics)	Stroh M, et al. (2010) 68	102 ng/mL h (7.5 mg)	0–inf	Yes	NA
Midazolam (i.v.)	Yes	Extensively	N05CD08 (Psycholeptics)	Pentikis HS, et al. (2007) 69	84.76 ng/mL h (2 mg; i.v. bolus)	0–inf	Yes	NA
Niraparib	Yes	Extensively	L01XX54 (Antineoplastic agents)	Moore K, et al. (2018) 70	29016.1 ng/mL h (300 mg, fasted)	0–inf	Yes	NA
Omeprazole	Yes	To lesser extent (mainly CYP2C19)	A02BC01 (Acid related disorders)	Ochoa D, et al. (2020) 71	2190.8 ± 2011.5 ng/mL h (40 mg, fasted)	0–inf	Yes	NA
Osimertinib	Yes	Primarily	L01EB04 (Antineoplastic agents)	Planchard D, et al. (2016) 72	2658 nM h (40 mg, 0–72 h); 5102 nM h (80 mg, 0–72 h); 15,480 nM h (160 mg, 0–72 h); 24,610 nM h (160 mg, 0–inf)	0–72 h or 0–inf	Yes	NA
Regorafenib	Yes	Primarily	L01EX05 (Antineoplastic agents)	Zhang Q, et al. (2021) 73	11354.7 ± 3323.9 ng/mL h (40 mg reference drug)	0–inf	Yes	NA
Ribociclib	Yes	Primarily	L01EF02 (Antineoplastic agents)	Ji Y, et al. (2020) 74	10,700 ng/mL h (600 mg)	0–inf	Yes	NA
Sorafenib	Yes	Primarily	L01EX02 (Antineoplastic agents)	Lathia C, et al. (2005) 75	11.04 mg/L h (50 mg)	0–inf	Yes	NA
Tivozanib	Yes	Partly	L01EK03 (Antineoplastic agents)	Cotreau MM, et al. (2015) 76	2223 ng/mL h (1.5 mg)	0–inf	Yes	NA
Triazolam	Yes	Primarily	N05CD05 (Psycholeptics)	Robin DW, et al. (1993) 77	15.57 ± 1.54 ng/mL h (0.25 mg)	0–inf	Yes	NA
Vandetanib	Yes	Partly	L01EX04 (Antineoplastic agents)	Martin P, et al. (2012) 78	22,030 ng/mL h (300 mg); 29,460 ng/mL h (400 mg); 61,140 ng/mL h (800 mg); 102,200 ng/mL h (1200 mg)	0–inf	Yes	NA
Vemurafenib	Yes	To lesser extend	L01XE15 (Antineoplastic agents)	Ribas A, et al. (2014) 79	119.0 ± 113.1 μg/mL h (960 mg, fasted)	0–inf	Yes	NA

Abbreviations: 0–inf, from zero to infinite hour; AUC, area under the plasma concentration–time curve; C _last, last observed concentrations; C _max, peak concentrations; i.v., intravenous; NA, not applicable; PK, pharmacokinetics.

^a Data extracted from concentrations–time curves by means of PlotDigitizer ¹⁰ and AUC_0–inf calculated using the trapezoidal rule.

Results of the analysis are presented in Figures 3, 4, 5 and Tables S1 and S2. Figure 3 presents all 53 human‐CYP3A4‐transgenic mouse AUC_0–inf and 19 wild‐type mouse AUC_0–inf in relation to the human AUC_0–inf after administration of an equivalent dose as absolute values and fold differences. Extrapolation with the exponent 0.67 resulted in a symmetric distribution of the fold differences in AUC_0–inf between human and both human‐CYP3A4‐transgenic and wild‐type mice around 1.13 and 1.02‐fold, respectively (where 1‐fold is an exact prediction of the AUC_0–inf; Figure 4b). Extrapolation with the exponent 0.75 resulted in a symmetric distribution around a median of 0.61 and 0.55‐fold for human‐CYP3A4‐transgenic and wild‐type mice, respectively. This suggests that using the allometric scaling exponent 0.67 results in more accurate predictions of the human equivalent dose than the exponent 0.75.

FIGURE 3.

(a) The AUC_0–inf of human‐CYP3A4‐transgenic and wild‐type mice plotted against the human AUC_0–inf after a human equivalent dose, (b) the distribution of the fold differences in AUC_0–inf between human and mice, and (c) the fold differences in AUC_0–inf between human and mice for each drug. Results from allometric scaling of the mice to human dose with the exponent 0.67 and 0.75 were both presented in all plots. Dotted lines in (a) represent the deviation from the line of unity (an ideal prediction of the human AUC_0–inf). Red and blue line represent the trend lines for wild‐type and human‐CYP3A4‐transgenic mice, respectively. All drugs were administered orally unless indicated differently. AUC_0–inf, area under the plasma concentration‐time curve until infinity.

FIGURE 4.

A selection of the publications which presented AUC_0–inf for both human‐CYP3A4‐transgenic and wild‐type mice (19 AUC_0–inf each): (a) the AUC_0–inf of human‐CYP3A4‐transgenic and wild‐type mice plotted against the human AUC_0–inf after a human equivalent dose, (b) the distribution of the fold differences in AUC_0–inf between human and mice, and (c) the fold differences in AUC_0–inf between human and mice for each drug. Results from allometric scaling of the mice to human dose with the exponent 0.67 and 0.75 were both presented in all plots. Dotted lines in (a) represent the deviation from the line of unity (an ideal prediction of the human AUC_0–inf). Red and blue line represent the trend lines for wild‐type and human‐CYP3A4‐transgenic mice, respectively. All drugs were administered orally unless indicated differently. AUC_0–inf, area under the plasma concentration‐time curve until infinity.

FIGURE 5.

Density plots of the fold differences in AUC_0–inf between mice and human for the exponents 0.67 and 0.75 for each study. To normalize fold differences to only values higher than one, one was divided by all the fold differences smaller than one. In addition, one was subtracted from all fold differences to normalize a perfect prediction (of 1‐fold) to zero. AUC_0–inf, area under the plasma concentration‐time curve until infinity; MAE, mean absolute error; RMSE, root mean squared error.

To perform a comparison between human‐CYP3A4‐transgenic and wild‐type mice for the predictability of the human equivalent dose, a selection was made of publications, which presented AUC_0–inf for both human‐CYP3A4‐transgenic and wild‐type mice (19 AUC_0–inf each; Figure 4). The selection resulted in a narrower interval between the first and third quartiles in the AUC_0–inf distributions for human‐CYP3A4‐transgenic compared to wild‐type mice (Figure 4b), suggesting that human‐CYP3A4‐transgenic mice are more accurate predictors of the human exposure than wild‐type mice for most drugs. However, allometric scaling with the exponent 0.67 resulted in a higher RMSE for human‐CYP3A4‐transgenic than wild‐type mice, 6.82 versus 4.96‐fold (normalized), respectively (Figure 5). For allometric scaling with the exponent 0.75, the RSME were 5.44 versus 5.10‐fold (normalized), respectively. The MAE was slightly lower for the human‐CYP3A4‐transgenic compared to wild‐type mice, 3.05 versus 3.08‐fold (normalized) for the exponent 0.67 and 3.06 versus 3.19‐fold (normalized) for the exponent 0.75, respectively. Removal of the extreme outlier ibrutinib resulted in results in favor of the human‐CYP3A4‐transgenic mice. Allometric scaling with the exponent 0.67 resulted in a lower RMSE for human‐CYP3A4‐transgenic than wild‐type mice, 2.88 versus 3.96‐fold (normalized), respectively (Figure 5). For allometric scaling with the exponent 0.75, the RSME were 4.49 versus 4.98‐fold (normalized), respectively. The MAE was also lower for the human‐CYP3A4‐transgenic compared to wild‐type mice, 1.72 versus 2.50‐fold (normalized) for the exponent 0.67 and 2.44 versus 2.99‐fold (normalized) for the exponent 0.75, respectively.

Large variability was observed between experiments with the same compound (ibrutinib and triazolam) and human‐CYP3A4‐transgenic mice (Table S1). Finally, between 57% and 79% of the predictions of both human‐CYP3A4‐transgenic and wild‐type mice fell within the toxicokinetic safety margin of four‐fold recommended by the World Health Organization to allow for interspecies differences and between 87% and 95% fell within the safety margin of 10‐fold for toxicokinetics and toxicodynamics combined, with no clear advantage for either mouse model (Table S2). ⁵

DISCUSSION

Human‐CYP3A4‐transgenic mouse for quantitative predictions in humans

Perhaps contrary to expectations, humanization of the mouse Cyp3a enzymes by means of knock‐out and replacement with the human CYP3A4 enzyme in general does not improve the predictions of exposure for CYP3A4‐metabolized drugs in humans. This result is mainly based on PK experiments in two mouse models developed by Hasegawa et al. and van Herwaarden et al. (Figure S1). Based on the RMSE, the human‐CYP3A4‐transgenic mouse model performs worse than the wild‐type mouse model. This is mainly caused by one extreme outlier, ibrutinib, in predictions of the human‐CYP3A4‐transgenic mouse model. Nevertheless, there is no obvious reason for excluding this drug because it is mainly metabolized by CYP3A4. ²² On the other hand, the percentage of predictions within fourfold and 10‐fold difference from the human exposure are slightly in favor of the human‐CYP3A4‐transgenic mouse model (Table S2). Everything considered, the small differences in predictability of human exposure suggest that the human‐CYP3A4‐transgenic mouse model will not markedly contribute to more accurate predictions of exposure following FIH doses in clinical trials by means of allometric scaling.

Difficulties in interspecies extrapolation

The CYP enzymes originate from a gene family that can be found in a wide range of organisms ranging from bacteria, plants, animals to humans, and even viruses. Over time, all species developed different variants of CYP enzymes with part having a common pivotal role, the detoxification of xenobiotics. It has been 75–125 million years ago that mice and humans had a common ancestor. The overlap in absorption, distribution, metabolism, and excretion (ADME) related processes probably stems from exposure to similar xenobiotics over this period, which resulted in similar evolutionary properties. This might explain why the wild‐type mice themselves already perform relatively well in the prediction of human exposure. Nevertheless, species have since then adapted to the exposure to partly different xenobiotics, resulting in deviation in detoxifying CYP enzymes to a greater or lesser extent, and the same applies for other ADME‐related processes. It is therefore important to elucidate what interspecies differences are responsible for these deviations in order to account for them in advance. Our hypothesis was that humanization of mouse CYP3A enzymes could be an important contributor to the reduction of the error in the predictions of human exposure. However, as it turns out, humanization of CYP3A enzymes alone is not nearly enough to account for the misspecifications in prediction of human exposure in the context of FIH dosing. Distinguishing two species by pinpointing one specific process (like CYP3A‐mediated metabolism) proves to be unrealistic. Interspecies differences consist of an interplay of many different processes that are vastly more complex, where the absence of a certain process in a species can be compensated by other processes. ²³ Here, we will discuss several examples of other interspecies differences that might contribute to the misspecifications observed in mice to human extrapolation.

First, the absorption of orally administered drugs is highly dependent on the biopharmaceutics classification system (BCS) class of a drug. The permeability and solubility define the class to which a drug is designated. However, the BCS classification in humans does not necessarily apply to mice. Solubility of drugs with a basic pK_a is different in the gastrointestinal tract of mice compared to humans. This is because the normal murine gastric pH is 3–4 and declines to an intestinal pH ~ 5, ²⁴ whereas the human gastric pH is 1–2.5, and intestinal pH is 6.5–7.5. ²⁵ As a result, absorption profiles differ because a drug can be fully protonated and ionized in the human stomach, whereas only partially in the mouse stomach. In addition, drugs are often administered as a solution by gavage into the stomach in mice, as opposed to solid dosing forms in humans.

Second, variation between experiments with the same drug and animal species is a topic that appears underexposed in the literature. It has been demonstrated that despite strict standardization of experiments, animals can behave differently between different laboratories ²⁶ and it is also been shown that mouse phenotypes can fluctuate, resulting in different results between batches. ²⁷ Consultation with preclinical scientists working with laboratory animals confirmed that deviations of about two‐fold in absolute drug concentrations between PK experiments of the same drug and mouse species over time (e.g., 6 months apart), even performed in one facility, is not uncommon. This is less of a concern if groups of animal strains within an experiment are treated equally and within a limited period of time, and only directly compared with each other. This results in minimal variability between groups apart from the investigated difference between the strains. In that context, continuity of absolute values (drug concentrations) over multiple experiments over a prolonged period of time is of lesser concern for answering certain hypotheses. However, for reliable quantitative predictions of the FIH dose, consistency over experiments and especially over time is warranted.

Third, replacing CYP3A enzymes in mice with human CYP3A enzymes does not necessarily imply that the corresponding overall metabolism will be similar. Expression and quantity of the replaced CYP3A enzymes can differ from that in humans, resulting in higher or lower clearance of the drug in the corresponding organ. For instance, van Herwaarden et al. suggests that CYP3A4 expression in the intestines of their transgenic strain is higher in mice compared to humans, which potentially results in lower bioavailability in mice. ⁷ Probably, there are many more physiological differences between mice and human that can influence drug PKs that still have to be elucidated. An example is a plasma protein expressed by mice, carboxylesterase 1c, which is absent in human plasma resulting in poor translation from mice to humans if the drug PK is influenced by carboxylesterase 1c through metabolism or strong binding. ²⁸ For example, cabazitaxel and everolimus are known to have high binding affinity to this protein resulting in PK differences between human and mouse. ²⁹ , ³⁰ In short, developing a mouse model that would suit all drugs would require many modifications with still a possibility of missing crucial ADME processes. More importantly, the trade‐off must be made as to whether the investments are worth the gains considering the already quite good performance of the wild‐type mice in terms of quantitative predictions (at least, for the panel of drugs considered in this analysis).

Other applications for the human‐CYP3A4‐transgenic mouse model

Despite the results of this meta‐analysis, the benefits of human‐CYP3A4‐transgenic mouse models for quantitative predictions of the human PKs could potentially be further exploited using modern data analysis approaches. More in depth knowledge of PKs in genetically modified animals can be obtained using PK modeling. In addition to finding a difference in exposure, PK modeling can uncover knowledge on the underlying PK processes that are potentially altered by drug metabolizing enzymes. The human‐CYP3A4‐transgenic mouse model will probably be more representative for the underlying PK processes for human CYP3A4. The human‐CYP3A4‐transgenic mouse model is therefore likely to be more accurate in the prediction of drug–drug interactions (DDIs) in which this enzyme plays a role. It allows a more evidence‐based approach for animal‐to‐human extrapolation. Two studies have applied population PKs and physiologically‐based PK (PBPK) modeling approaches to analyze quantitative results generated in human‐CYP3A4‐transgenic mouse models in order to predict human exposure. ²⁰ , ²¹ We described the extrapolation of four compounds in human‐CYP3A4‐transgenic mouse models to humans using a population PK approach. The use of a population PK approach enabled the authors to correct for species differences they assumed to be relevant for the compounds concerned, resulting in more accurate predictions of the human exposure with human‐CYP3A4‐transgenic compared to wild‐type mouse models. Zhang et al. used a PBPK modeling approach in combination with a boosting effect study of ritonavir on NVS123 in a human‐CYP3A4‐transgenic mouse model. Hereby, not only the human exposure could be accurately predicted, but a DDI involving CYP3A4 was described as well. Choo et al. reported that under specific conditions the transgenic mouse model may be a useful tool to predict the relative contribution of hepatic and intestinal metabolism. They anticipate that in future PBPK modeling in combination with in vitro data will help to clarify the utility and limitations of the transgenic models. To summarize, PK modeling approaches can help to correct for interspecies differences that are expected to contribute to deviations in the predictions. Nevertheless, this requires prior knowledge of interspecies differences (e.g., differences in gastrointestinal tract pH, enzyme or transporter expression, or binding partners proteins) and therefore FIH dose predictions remain difficult. Human‐CYP3A4‐transgenic mouse models, extended with modeling approaches, are therefore probably best suited to more accurately predict CYP3A4 inhibition and induction DDIs in a quantitative way, for compounds for which there is already some clinical PK data is available to correct the human‐CYP3A4‐transgenic mouse models extrapolation to humans and validate predictions.

AUTHOR CONTRIBUTIONS

D.D., A.H.S., J.H.B., A.D.R.H., and T.P.C.D wrote the manuscript. D.D., J.H.B., T.P.C.D., and A.D.R.H. designed the research. D.D. performed the research. D.D. analyzed the data.

FUNDING INFORMATION

No funding was received for this work.

CONFLICT OF INTEREST STATEMENT

The authors declared no competing interests for this work.

Supporting information

Figure S1

Damoiseaux D, Schinkel AH, Beijnen JH, Huitema ADR, Dorlo TPC. Predictability of human exposure by human‐CYP3A4‐transgenic mouse models: A meta‐analysis. Clin Transl Sci. 2024;17:e13668. doi: 10.1111/cts.13668

DATA AVAILABILITY STATEMENT

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.