Drug Interaction PBPK Modeling: Review of the Literature Exposes the Need for Increased Verification of Model Inputs and Outputs as Part of Credibility Assessment

ABSTRACT

In vitro data are routinely used to support both static and physiologically based pharmacokinetic (PBPK) model‐based drug–drug interaction (DDI) predictions. Such efforts are possible after years of rapid progress, enabled by the greater availability of in vitro reagents, kinetic models, and ready access to PBPK software packages with increased computing power supported by drug‐specific compound files. While acknowledging the progress, however, various investigators have documented the challenges and pitfalls associated with PBPK modeling and have called for improved model verification, credibility assessment, and greater confidence building. As summarized in the current narrative, a review of the DDI literature does expose the need for PBPK model parameter (input and output) verification. Representative examples of PBPK‐based modeling involving induction (cytochrome P450 (CYP) 3A4 and organic anion transporting polypeptide 1B1 and 1B3), pregnancy‐associated upregulation (CYP2D6), and inhibition (CYP1A2‐mediated metabolism and creatinine renal clearance) are described. The narrative also includes the clinical application of biomarkers (e.g., CYP3A4 and CYP2D6) and tissue biopsy expression profiling as a means of providing additional mechanistic information and DDI data that are independent and complementary to PBPK models. With the advent of in vitro microphysiological systems, biomarkers, burgeoning plasma‐based (liquid) biopsy protocols, and the possibility of machine learning‐enabled literature searches integrated with modeling software, it is envisioned that such tools could be used jointly to further enhance PBPK model verification efforts within a predefined credibility assessment framework. Ultimately, the goal is to deploy PBPK modeling with greater confidence in lieu of time and resource‐intensive clinical DDI studies.

1. Introduction

It is accepted that pharmacokinetic (PK) modeling and simulation, especially physiologically based pharmacokinetic (PBPK) modeling, can enable most facets of the drug discovery and development process. As reviewed extensively by various investigators across the pharmaceutical industry, inclusion of PBPK modeling in submissions by sponsors has increased. This is likely because PBPK models can support human dose and PK projections prior to first in man. They can also facilitate dose selection and optimization, formulation development, PK‐pharmacodynamics analysis, and enable the modeling of study drug PK in individuals across lifespan and with variable physiology (e.g., age, body weight, pregnancy status, and organ function), disease status (e.g., cancer, infection, inflammation), drug‐metabolizing enzyme genotype (phenotype), and drug transporter genotype (phenotype) [1, 2, 3, 4, 5, 6, 7]. When it comes to PK modeling, the study of drug–drug interactions (DDI) involving induction, inhibition, and dual mechanisms (induction plus inhibition) has also received considerable attention [8, 9, 10, 11, 12, 13, 14]. For example, at the time of writing an online literature search (Pubmed) retrieved 946 entries (1989 to 2025) using the search phrase “Drug‐drug interaction PBPK” (~90% and 70% of the entries over the time period of 2011 to 2025 and 2018 to 2025, respectively).

Most sponsors now conduct in vitro studies to profile their new molecular entities as DDI perpetrators (inhibitors, inducers, precipitants) and victims (substrates, objects). Such in vitro data are used as input for various simple static, expanded static (“net effect”, Figure S1) and PBPK‐based DDI prediction models in the hope of de‐risking DDI (Figure 1). In the case of DDI victims, the fraction and rate of metabolism by the impacted enzyme (f _m), or rate and fraction transported by the impacted transporter (f _t), can be derived in vitro and in vivo (e.g., human radiolabel study or clinical DDI study). For DDI perpetrators, in vitro input parameters like IC₅₀ (concentration of inhibitor rendering a 50% reduction in activity) and K _i (inhibition constant) describe reversible inhibition; k _inact (maximal rate of enzyme inactivation) and K _I (inhibitor concentration at half of k _inact) are used for mechanism‐based inhibition; and E _max (maximal induction) and EC₅₀ (half‐maximal induction) describe induction (Figure 1). Ideally, the output of such modeling exercises is assessed (e.g., bias, precision, and error) and is considered of sufficient quality to enable the design of DDI studies and drive informed decision making (e.g., prioritization, de‐prioritization, or need for a formal DDI study) [8, 9, 10, 11, 12, 13, 14]. Not surprisingly, the topic of PBPK modeling has also been the subject of various regulatory agency guidance documents, which now include the recently issued ICH (International Council for Harmonization of Technical Requirements for Pharmaceuticals for Human Use) M12 DDI guidance [15, 16]. However, although PBPK modeling is now widespread and routine, various investigators have described the challenges and pitfalls, the need for greater confidence in the models, and some have called for improved model verification and the establishment of credibility assessment frameworks [17, 18, 19, 20, 21, 22].

FIGURE 1.

High‐level summary of model‐based predictions of victim and perpetrator DDI. V _max, maximal rate of enzyme reaction; K _m, concentration of substrate at half‐maximal rate of enzyme reaction or transport; J _max, maximal rate of transporter‐mediated uptake or efflux; f _m, fraction metabolized by the inhibited enzyme; f _t, fraction transported by the inhibited transporter; IC₅₀, concentration of inhibitor rendering a 50% reduction in activity; K _i, inhibition constant; k _inact, maximal rate of enzyme inactivation; K _I, inhibitor concentration at half of k _inact; E _max, maximal induction; EC₅₀, half‐maximal induction; P_app, apparent permeability; f _u,plasma, fraction unbound in plasma; f _u,inc, fraction unbound in the in vitro incubation; PBPK, physiologically based pharmacokinetics; AUC, area under the plasma concentration versus time curve; AUC ratio, AUC_{(with perpetrator)}/AUC_{(control, reference, or placebo)}; C_max, maximal plasma concentration; C _max ratio, C _{max(with perpetrator)}/C _{max(control, reference, or placebo)}; t1/2, plasma half‐life; t1/2 ratio, t1/2_{(with perpetrator)}/t1/2_{(control, reference, or placebo)}; CL_renal, renal clearance; CL_renal ratio, CL_{renal(with perpetrator)}/CL_{renal(control, reference, or placebo)}; AFE, average fold error; GMFE, geometric mean fold error; RMSE, root mean square error; % prediction error = [(predicted value—observed value)/observed value] × 100.

After a careful review of the DDI modeling literature, it is apparent that independent verification of model input and output parameters is needed as part of a credibility assessment framework. Five representative examples are described in the following narrative involving the modeling of hepatic organic anion transporting polypeptide (OATP) 1B1 and 1B3 (OATP1B1/3) induction by rifampicin, inhibition of creatinine renal secretory clearance mediated by organic cation transporter 2 (OCT2) versus organic anion transporter 2 (OAT2), induction of cytochrome P450 (CYP) 3A4 (CYP3A4) in gut versus the liver, inhibition of CYP1A2 by oral contraceptives containing 17α‐ethinyl estradiol, and pregnancy‐related changes in liver CYP2D6 expression and activity. From the outset, it should be noted that all the authors are commended for their efforts, especially when some lacked mechanistic information, that they were careful to conduct sensitivity analyses, reported the precision and bias of their models, and provided tabular summaries of input parameters and data sources more or less in compliance with agency guidance [16]. Ultimately, a goal of this review is to open scientific dialog and discourse on parameterization of PBPK models ensuring consistency with existing literature to help build credibility in the modeling output. In addition to the five examples, the integration of modeling with biomarkers, selective for drug‐metabolizing enzymes and drug transporters, and human tissue biopsy data is also described.

2. Biomarkers and Tissue Biopsy

2.1. Tissue Biopsy

As has been reviewed previously, there are numerous examples in the literature describing the expression profiling of drug‐metabolizing enzymes, transporters, transcription factors, and various nuclear hormone receptors in gut and liver tissue biopsy samples [23, 24]. Because of the invasive surgical procedures required to obtain organ tissue biopsies, and the potential for complications related to such procedures, there are only a small number of studies documenting dual intestine and liver biopsy profiling. Some examples of studies reporting tissue biopsy data are described in Table S1, wherein various authors evaluated the induction of CYP3A4 following agents such as efavirenz, rifampicin, and carbamazepine. Likewise, ex vivo inhibition of CYP3A4 activity in gut biopsy samples has been reported following diltiazem and clarithromycin administration (see references in Table S1). In most cases, investigators were able to relate tissue biopsy profiling data to various urine‐ and plasma‐based trait measures such as biomarker metabolite‐to‐parent ratios and the PK of intravenous (versus oral) dosed probe drugs like midazolam, erythromycin, and felodipine.

As shown in Figure 2, profiling of tissue biopsy samples pre‐ versus post‐perpetrator dosing, although invasive, has the advantage of providing a direct readout of the enzyme expression fold‐increase following an inducer (Figure 2 term B ≥ 1 and equal to the expression fold‐increase in the liver; term Y ≥ 1 and equal to the expression fold‐increase in the gut) or the decrease in ex vivo activity following a mechanism‐based inhibitor (term A ≤ 1 and equal to the fraction activity remaining in the liver; term X ≤ 1 and equal to the fraction activity remaining in the gut) at the time of sample collection. This means that investigators are not required to use in vitro‐derived parameters (e.g., K _I, k _inact, EC₅₀, and E _max) or challenging estimates of the perpetrator exposure at the site of the DDI. In addition, it is possible to avoid problematic assumptions regarding the turnover of enzymes, such as CYP3A4 in the gut and liver, and circumvent the use of complex scaling factors, the use of calibrator perpetrators, adjustment of in vitro parameters to reclaim the observed DDI, and use of elaborate in vitro to in vivo extrapolations (Figure S1). Of note, profiling of ribonucleic acid (mRNA) and protein expression in biopsy samples is also applicable to the study of enzyme and transporter expression decreases resulting from suppression by cytokines, transcription factors, and nuclear hormone receptors (term B ≤ 1 and term Y ≤ 1). Furthermore, the “net effect” of induction and mechanism‐based inhibition can be determined by measuring ex vivo enzyme activity in tissue biopsy samples, although the additional measurement of protein expression is advisable. Unfortunately, it is not possible to study reversible inhibition using biopsy samples because of perpetrator drug washout during sample handling.

FIGURE 2.

Clinical study of perpetrator DDI using biomarkers and tissue biopsy. sEV, small extracellular vesicles; f _g, fraction surviving gut first pass; f _m, fraction metabolized by target enzyme; X, fraction activity remaining in enterocytes; A, fraction activity remaining in liver; Y, expression fold‐increase in enterocytes; B, expression fold‐increase in the liver; AUC, area under the plasma concentration versus time curve (AUC) ratio (AUC_+perpetrator/AUC_baseline); 6βHC, 6β‐hydroxycortisol; 4βHC, 4β‐hydroxycholesterol; 1βHDCA, 1β‐hydroxydeoxycholic acid; OATP, organic anion transporting polypeptide; CYP3A4, cytochrome P450 3A4; PK, pharmacokinetics; PBPK, physiologically based pharmacokinetics. See Figure S1 for more details.

Obtaining tissue biopsies from consenting healthy volunteers is not straightforward and requires rigorous institution review board oversight and approval that is consistent with regulatory guidance [25]. Therefore, some investigators have focused their attention on the isolation and profiling of plasma‐ and serum‐derived small extracellular vesicles (sEV) as liquid biopsy, since they are less invasive and mostly retain the transcriptomic, proteomic, and metabolomic signature of their tissue of origin [23, 24, 26, 27, 28, 29, 30]. To date, modafinil [weak to moderate inducer based on midazolam AUC ratio of 0.65 (once daily 200 mg × 7 days) and a triazolam AUC ratio of 0.42 (once daily 200 mg × 7 days followed by 400 mg × 21 days), Table S2], rifampicin [strong inducer based on midazolam AUC ratio < 0.2 (once daily 600 mg × 14 days), Table S2], and PF‐06835919 are the only three compounds tested as inducers after CYP3A4 expression profiling of plasma‐derived liver and non‐liver (global minus liver) sEV. In all three cases, it was possible to determine the fold increase in sEV CYP3A4 expression and predict the area under the plasma concentration versus time curve (AUC) ratio (AUC_inducer/AUC_reference) for oral midazolam [26, 27, 28]. The report by Qui et al., who clinically assessed a Pfizer development candidate (PF‐06835919) as a CYP3A4 inducer because of positive in vitro human hepatocyte data, is highly illustrative of the challenges faced when predicting DDI [28]. In this instance, the authors had no CYP3A4 biomarker data in hand, had obtained static model predicted AUC ratios ranging from 0.23 to 0.79, and the observed midazolam AUC ratio was close to unity (0.97). Subsequent profiling of plasma‐derived sEV revealed no induction of CYP3A4 protein expression or enzyme activity, and a more reasonably predicted midazolam AUC ratio of 0.9 (enzyme activity) to 1.0 (protein expression). Therefore, as in the case of tissue biopsy data, plasma‐derived sEV data can also serve as input for static models to predict induction without knowledge of the inducer concentration in the liver and intestine and its induction parameters E _max and EC₅₀ (Figure 2).

Although plasma sEV isolation protocols are not well established, require validation, and only limited data are available, CYP3A4 expression fold‐increases in plasma‐derived non‐liver (gut) and liver sEV are consistent with conventional tissue biopsy data for the model inducer rifampicin (Table S1). Most importantly, the cost and potential medical complications that can arise from conventional invasive biopsy procedures will hinder the routine use of tissue biopsies in support of clinical pharmacology DDI studies. Alternatively, plasma‐derived tissue sEV only require a blood draw, and once profiled, the data can be integrated with plasma biomarker and drug probe PK trait measures. Such attributes support the continued development of plasma sEV‐based liquid biopsy isolation and bioanalytical methods.

2.2. Biomarkers

Alongside emergent liquid biopsy approaches, as well as innovations such as drug probe cocktails and microdosing, investigators have also leveraged plasma‐ and urine‐based biomarkers selective for different drug‐metabolizing enzymes and transporters (Figure 2) [24]. Some examples include plasma 4β‐hydroxycholesterol (versus cholesterol) and 1β‐hydroxydeoxycholic acid for CYP3A4, urine and/or plasma 6β‐hydroxycortisol (formation clearance and 6β‐hydroxycortisol‐to‐cortisol ratio) and a solanidine metabolite (3,4‐seco‐solanidine‐3,4‐dioic acid) for CYP3A4 and CYP2D6, respectively, and plasma coproporphyrin I for OATP1B1/3 [24, 31, 32, 33, 34, 35, 36]. This means that one can study the induction of CYP3A4, or CYP2D6 and OATP1B1/3 phenotype and inhibition, without having to administer a probe drug in a Phase I clinical trial. When compared to drug probes, biomarkers can also facilitate the dose‐dependent assessment of DDI and tracking of DDI time courses (DDI onset versus washout).

Examples of various CYP3A4 inducers studied with plasma 4β‐hydroxycholesterol are described in Tables S2–S4. However, some have questioned the utility of 4β‐hydroxycholesterol as a biomarker because its ratio (versus cholesterol or 4α‐hydroxycholesterol) in plasma is thought to be attributed mostly to liver CYP3A4 [37, 38, 39, 40, 41]. This is reflected in the various PK models that have been described for 4β‐hydroxycholesterol [42, 43, 44]. Evidently, the lack of effect of grapefruit juice on plasma 4β‐hydroxycholesterol and the weaker correlation of plasma 4β‐hydroxycholesterol with gut (versus liver) tissue biopsy and sEV CYP3A4 expression (activity) support such a conclusion (Table S1) [27, 38, 41]. However, cholesterol can be metabolized in the intestine, and some investigators have hypothesized that plasma 4β‐hydroxycholesterol likely reflects hepatic CYP3A4 in the basal state and that both the gut and liver contribute after administration of inducers known to increase protein expression in both organs [27, 37, 39]. Of note, the long apparent half‐life of 4β‐hydroxycholesterol (~17 days) makes it unsuitable to study the acute effects of induction or inhibition in humans. Additionally, the relatively low dynamic signal to detect changes in 4β‐hydroxycholesterol with moderate–mild CYP3A induction DDI makes it less attractive as a routine biomarker to use in clinical development [35].

More recently, 1β‐hydroxydeoxycholic acid, along with its glycine and taurine conjugates, has been proposed as an alternative CYP3A4 biomarker. Because of its shorter plasma half‐life versus 4β‐hydroxycholesterol (4 versus 17 days), changes in plasma 1β‐hydroxydeoxycholic acid appear more dynamic for both CYP3A4 inhibitors and inducers compared to 4β‐hydroxycholesterol (Tables S2–S4) [35]. However, the role of gut versus liver CYP3A4 in governing plasma 1β‐hydroxydeoxycholic acid is unknown, but it is assumed that parent deoxycholic acid can also undergo CYP3A4‐catalyzed 1β‐hydroxylation in enterocytes and that there is very extensive conjugation with taurine and glycine catalyzed by liver‐enriched bile acid‐CoA: amino acid N‐acyltransferase [45, 46, 47]. Even with these caveats, the early data using 1β‐hydroxydeoxycholic acid to detect both inhibition and induction DDI in humans are very encouraging and warrant further study.

In the context of DDI predictions, it is likely that investigators will increasingly blend biomarker and liquid biopsy data with their PK modeling. For example, a PBPK model has been described for 4β‐hydroxycholesterol, and plasma sEV data have been used to predict changes in plasma 4β‐hydroxycholesterol following administration of an inducer [27, 42]. Therefore, early Phase I clinical study biomarker and liquid biopsy data could be used in a complementary manner to validate a priori and subsequent PBPK modeling exercises, helping to establish credibility for new molecular entities when CYP3A4 induction is suspected. In the case of plasma 1β‐hydroxydeoxycholic acid, its greater dynamic range for inhibitors could support its use to study mechanism‐based inhibition alongside plasma‐derived sEV ex vivo CYP3A4 activity measurements.

3. Examples of Integrating DDI Modeling With Biopsy Data

As described previously, there are numerous literature reports describing perpetrator and victim DDI modeling and simulation exercises and most have focused on a comparison of the observed versus predicted result, the prediction error, as well as model bias and precision (Figure 1) [9, 10, 11, 13, 14]. However, very few investigators have attempted to integrate their modeling outputs with independent information such as published tissue biopsy data. In the following, two examples are described in some detail to illustrate the usefulness of an integrative approach.

3.1. Induction of Intestinal P‐Glycoprotein by Rifampicin

In the first example, Yamazaki et al. modeled the induction of intestinal P‐glycoprotein by rifampicin and its impact on the PK of four drugs (digoxin, talinolol, quinidine, and dabigatran) [48]. For each of the drugs, a sensitivity analysis was described and involved relating the % prediction error for the predicted victim drug maximal plasma concentration ratio and AUC ratio [% prediction error = [(predicted value—observed value)/observed value] × 100] to the extent of intestinal P‐glycoprotein induction (1‐ to 10‐fold). In all four cases, the authors reported that the prediction error was low (≤ 25%) when the fold increase in gut P‐glycoprotein expression ranged from 2 to 4, consistent with independently published tissue biopsy data [23, 48].

3.2. Inhibition of Intestinal and Hepatic CYP3A4 by Clarithromycin

In the second case, Quinney et al. studied the inhibition of intestinal and hepatic CYP3A4 by clarithromycin (500 mg twice daily for a week) and developed a semi‐PBPK model to describe the time‐dependent decrease in enzyme activity in both organs [49]. Small bowel biopsies were obtained via endoscopy at baseline and on three additional occasions post‐clarithromycin. The derived tissue was subject to immunoblotting (CYP3A4 expression) and CYP3A4 ex vivo enzyme activity measurement (1′‐hydroxymidazolam formation from midazolam). In the absence of liver biopsy, hepatic CYP3A4 activity was determined in each of their study subjects after administration of intravenous midazolam and reporting of a 3‐h serum 1′‐hydroxymidazolam‐to‐midazolam ratio. Although there was considerable variation, there was no significant change in gut biopsy CYP3A4 protein expression while a significant (p < 0.05) mean 64% decrease (versus pre‐dose) in ex vivo CYP3A4 activity was reported in the same samples. The authors also modeled the clarithromycin concentration versus time profiles in the liver and gut wall and predicted the changes in CYP3A4 activity therein. Importantly, they also successfully integrated the model‐based prediction with the gut biopsy activity and serum 1′‐hydroxymidazolam‐to‐midazolam ratio data and concluded that gut CYP3A4 was more inhibited compared to liver CYP3A4.

3.3. Challenges

Although the two examples above illustrate the usefulness of integrating biopsy data with PK modeling, some challenges do remain. For example, rifampicin can induce numerous drug‐metabolizing enzymes and is known to impact zidovudine PK [50]. As a DDI victim, zidovudine is largely metabolized by UDP‐glucuronosyltransferases (UGT) expressed in the gut (UGT2B7) and liver (UGT2B4 and UGT2B7), with some metabolism to a 3′‐amino metabolite via CYP3A4 and other CYPs [50, 51, 52, 53]. Therefore, induction of both UGTs and CYP3A4 would be expected to explain the observed AUC ratio (0.54) for parent zidovudine following rifampicin, as well as the two‐fold increase in formation clearance of both the zidovudine glucuronide and 3′‐amino metabolite. However, mRNA profiling of both gut and liver biopsy samples after rifampicin dosing presents no induction of UGT2B4 or UGT2B7 mRNA even when CYP3A4 mRNA is induced [51, 52]. This implies that in vitro‐based predictions of the rifampicin‐zidovudine DDI would be challenging and that PBPK modeling would not be straight forward [53]. One possibility that may explain such a disconnect is the rifampicin‐mediated modulation of various microRNA (miR) species that would give rise to upregulated UGT2B4 and UGT2B7 protein expression, and hence zidovudine glucuronidation, without impacting mRNA levels [54]. Testing of such a hypothesis would require rifampicin dosing, harvesting of tissue or liquid biopsy samples, and integrated expression profiling of miR and UGT mRNA and protein expression. Such studies are worthwhile because different investigators have reported that rifampin also reduces the plasma concentrations of oral and intravenous hydromorphone, accompanied by increases in its hydromorphone glucuronide‐to‐hydromorphone ratio, which further implicates induction of both gut and liver UGT2B7 activity [55]. Of note, static modeling of the observed rifampicin‐zidovudine interaction indicates that, along with CYP3A4 induction, one only requires ~1.7‐fold increase in liver UGT activity to render an AUC ratio of 0.54 (Figure S2).

4. Examples of DDI Modeling

4.1. Modeling of CYP3A4 Induction in the Intestine and Liver

Expression of drug‐metabolizing enzymes like CYP3A4 and CYP3A5, in both the gut wall and the liver, is governed by numerous transcription factors and nuclear hormone receptors such as the pregnane X receptor, growth hormone, glucocorticoid receptor, and constitutive androstane receptor. In enterocytes, the vitamin D receptor can also play a role in regulating CYP3A4 expression. Therefore, a given inducer can manifest a unique induction signature across both organs, governed by its differential interaction with different receptors (E _max/EC₅₀ ratio and full versus partial agonism profile) as well as its exposure and residence time in each organ. In addition, a subject's genotype (e.g., P‐glycoprotein and vitamin D receptor) can also impact basal expression (e.g., CYP3A4) and hence determine the fold changes with the inducer [56, 57, 58]. Therefore, induction of an enzyme like CYP3A4 in the intestinal tract does not mean that hepatic CYP3A4 will be induced and vice versa. Such considerations are important when modeling DDI involving induction, especially if human primary hepatocyte‐derived E _max and EC₅₀ values are used to predict induction in the intestine.

To date, available tissue biopsy data are supportive of the notion that inducers present varied CYP3A4 induction profiles when it comes to the gut versus the liver (Figure 3, Table S1). For carbamazepine, available tissue biopsy data indicate ~2 to 4‐fold induction of gut and liver CYP3A4. Biopsy data present no induction of gut CYP3A4 by efavirenz accompanied by 1.3‐ to 1.8‐fold increases in plasma 4β‐hydroxycholesterol. Presumably, the latter reflects induction of liver CYP3A4 only. Across different studies, tissue biopsy profiles of subjects dosed with rifampicin (600 mg) point to a ~3‐to 4‐fold increase in liver CYP3A4. Although induction of intestinal CYP3A4 appears more varied (~2‐ to 10‐fold), the largest numbers of studies reporting gut biopsy data present fold increases in the range of 2 to 4 (Figure 3B). In agreement, more recent plasma‐derived sEV liquid biopsy data present 3.7‐ and 4.4‐fold increases in CYP3A4 protein expression in liver and non‐liver (gut), respectively, following rifampicin (600 mg for 14 days) [26]. By comparison, modafinil, a moderate inducer based on its impact on oral midazolam AUC (Figure 3C), induces non‐liver (gut) and liver sEV CYP3A4 expression 1.9‐ and 1.3‐fold, respectively (Table S1) [27]. Of note, induction of polymorphic CYP3A5 has also been studied after isolation of plasma‐derived liver sEV of rifampicin‐dosed CYP3A5 expressing subjects (600 mg × 14 days) and found to be less induced versus CYP3A4 [N = 4 subjects; mean fold‐increase (90% confidence interval) = 3.2 (1.0 to 6.0) vs. 1.3 (0.8 to 1.5); p = 0.039] [26].

FIGURE 3.

Evaluation and prediction of CYP3A4 induction based on tissue and liquid biopsy profiling and oral midazolam plasma AUC ratio. (A) Relating CYP3A4 expression fold‐increase in gut (Y, filled triangles) and liver (B, filled circles) to the predicted oral midazolam AUC ratio (bars), assuming f _m = 0.9; f _g = 0.51 [26, 27]. The equation shown has been simplified and only considers induction (based on equation in Figure 2). For modafinil (Y > B), rifampicin (B = Y) and carbamazepine (B = Y), the shaded area represents the reported range of CYP3A4 increases in biopsy samples and the reported oral midazolam AUC ratio (references in Tables S1 and S2). Only model data for efavirenz are shown in the figure (assuming B > Y). (B) Fold‐increase in CYP3A4 expression (or activity) in tissue and liquid (plasma‐derived sEV) biopsy samples following induction. References for source data are described in Table S1. Modafinil liquid biopsy data were obtained following once daily 400 mg × 14 days. (C) Reported oral midazolam plasma AUC ratios following different inducers (references for source data described in Table S2). AUC ratio for midazolam following once daily modafinil 200 mg × 7 days is shown. AUC ratio (AUCR), area under the plasma concentration versus time curve (AUC) ratio (AUC_inducer/AUC_reference); sEV, small extracellular vesicles; CYP3A4, cytochrome P450 3A4; f _g, fraction surviving gut first pass; f _m, fraction metabolized by impacted enzyme (CYP3A4); MOD, modafinil; RIF, rifampicin; CBZ, carbamazepine; (M), mRNA; (P), protein; (EVA), ex vivo activity; mRNA, messenger RNA.

From the standpoint of PBPK modeling, CYP3A4 induction by rifampicin is well represented in the literature, although investigators have had to address autoinduction and possible induction by its metabolites and have not considered CYP3A5. However, due to the identification of potentially genotoxic nitrosamine impurities in rifampicin formulations, sponsors of new molecular entities have sought to switch to alternatives such as phenytoin and carbamazapine [59, 60]. Therefore, induction of CYP3A4 in the intestine versus liver by different agents has garnered the attention of various investigators [59, 61, 62, 63, 64, 65, 66]. This is important because the majority of published PBPK modeling exercises have related to victims such as oral midazolam, which undergoes CYP3A4‐mediated metabolism in the intestine and liver. Of note, most investigators have described the performance of their models, reported model bias and precision, and conducted sensitivity analysis, but most have not related their model outputs to independent data sources such as tissue biopsy. Although most of them have described successful modeling of induction following rifampicin (e.g., 600 mg QD × 14 days), and its impact on victim drug PK, it is noted that fold‐increases for gut CYP3A4 have ranged from 3 to 10 [59, 62, 63, 64, 65, 66]. For example, Asaumi et al. reported 3‐fold and 3.5‐fold increases in gut and liver CYP3A4 following rifampicin, respectively [65]. More recently, Chen and Jones reported 10‐fold and 5.5‐fold increases in gut and liver CYP3A4, respectively [59]. Evidently, some of these model‐derived fold‐increases differ markedly from published biopsy data (Figure 3B, Table S1). Similarly, successful PBPK‐based modeling of induction by carbamazepine has consistently presented a ~2‐fold increase in liver CYP3A4 in agreement with available biopsy data (Figure 3B, Table S1) [59, 61]. However, published PBPK models for carbamazepine present more varied induction of CYP3A4 in the intestine, ranging from 2.5‐fold to ~5 to 8‐fold [59, 61]. This has forced some authors to explain their differentiated model outputs when compared to earlier reports [61]. With the unfortunate need to switch away from rifampicin for clinical pharmacology DDI studies, and limited tissue biopsy data, it is hypothesized that liquid biopsy data (e.g., CYP3A4 expression profiling of plasma‐derived gut and liver sEV) following the newly proposed step up (100–300 mg) carbamazepine dosing regimen (Table S2) would enable validation of the relative changes in CYP3A4 in the intestine and liver used for predictive PBPK models. The same is true for other rifampicin substitutes such as phenytoin.

As described in Figure 3A, one does not need to evoke fold‐increases in CYP3A4 expression > 5‐fold to generate midazolam AUC ratios consistent with clinical data (Figure 3C, Table S2). In fact, available tissue and liquid biopsy data can support reasonable AUC ratio predictions (Figure 3B, Table S1). For rifampicin, 3‐ to 4‐fold increases in liver and intestine CYP3A4 can render a midazolam AUC ratio < 0.2. Modafinil can bring about the observed midazolam AUC ratio (0.65) with a ~2‐fold and ~1.3‐fold increase in gut and liver CYP3A4, respectively. Likewise, a 2‐ to 3‐fold increase in both gut and liver CYP3A4 can explain the reported midazolam AUC ratio (~0.3) following carbamazepine, although inter‐subject variability is expected regarding fold‐increases in each organ. In contrast, efavirenz presents no induction of gut CYP3A4 (Table S1) and one would need to evoke a ~2‐fold increase in liver CYP3A4 activity, which theoretically would give rise to a midazolam AUC ratio of ~0.5. The same analysis can be extended to CYP3A4 biomarkers like 4β‐hydroxycholesterol and 1β‐hydroxydeoxycholic acid (Figure 4). In this instance, the fold‐increases in CYP3A4 expression described above can also explain plasma biomarker increases following carbamazepine (chronic dosing in epileptics and a step up dosing clinical pharmacology DDI induction protocol), rifampicin, modafinil, and efavirenz, with reasonable assumptions regarding the plasma half‐life of each biomarker and the turnover of CYP3A4 in gut and liver (Tables S2–S4, Figure S3) [27]. Interestingly, only by evoking induction in the gut and liver by rifampicin can one obtain reasonable predictions of plasma 4β‐hydroxycholesterol and 1β‐hydroxydeoxycholic acid increases (Table S3). This supports the hypothesis that both biomarkers may be dominated by liver CYP3A4 in the basal state, but both gut and liver can contribute after induction. The present analysis also supports our belief that consistently leveraging existing tissue and liquid biopsy data, both proprietary and taken from the literature, for commonly used CYP3A4 clinical index inducers will strengthen the credibility of the PBPK models for predictive purposes.

FIGURE 4.

Relating CYP3A4 fold‐increase in gut and liver to fold‐increase in plasma 4β‐hydroxycholesterol and 1β‐hydroxydeoxycholic acid. Fold‐increases in gut (shaded bars) and liver (unshaded bars) CYP3A4 were used as model input to predict the fold‐increase plasma 4β‐hydroxycholesterol (4βHC) and 1β‐hydroxydeoxycholic acid (1βHC) following carbamazepine assuming induction in gut ~ liver (A), rifampicin assuming induction in gut ~ liver (B), modafinil induction in gut > liver (C), and efavirenz induction in liver > gut (D). A model for 4β‐hydroxycholesterol has been published and was adapted for 1β‐hydroxydeoxycholic (Table S3) [27]. For each inducer, the shaded area represents the reported range of CYP3A4 fold‐increases in gut and liver biopsy (tissue and plasma‐based sEV) samples and the published fold‐increase in plasma 4β‐hydroxycholesterol and 1β‐hydroxydeoxycholic acid (see references in Tables S1 and S2). sEV, small extracellular vesicles.

4.2. Modeling the OCT2‐ Versus OAT2‐Mediated Renal Clearance of Creatinine

Creatinine renal clearance is well studied, as it has been widely used to estimate glomerular filtration rate as a measure of kidney function. Although monitoring of serum creatinine levels in clinical studies is common practice, it is now recognized that creatinine also undergoes active secretion mediated by renal transporters such as basolateral OCT2 and apical multidrug and toxin extrusion protein (MATE) 1 and 2‐K (MATE1 and MATE2‐K) [67, 68]. Therefore, any new molecular entity, or co‐administered drug, that inhibits one or more of these transporters can impact serum creatinine readouts, interfering with renal toxicity assessment in the absence of other functional markers such as cystatin‐C [68, 69].

Beyond OCT2, various groups have proposed that basolateral OAT2 also plays a role in creatinine renal secretory clearance [70, 71]. However, most investigators have relied on transfects of OAT2 in various cell lines (e.g., human embryonic kidney 293 cells and Madin–Darby canine kidney cells) because of the lack of readily available primary human renal proximal tubule epithelial cells (RPTEC) manifesting suitable OAT2 versus OCT2 functionality. More recently, however, Mathialagan et al. confirmed that transfected OAT2 and OCT2 can transport creatinine and were able to show that the former plays no role in creatinine uptake by freshly prepared (multi‐donor) human RPTEC preparations [72]. The authors were careful to demonstrate that OAT2 was active in the RPTEC preparations by using a substrate (penciclovir) and inhibitor (indomethacin) highly selective for OAT2 versus OCT2. The same RPTEC cell model also presented metformin uptake (OCT2/MATE mediated versus OAT2) that was inhibited by decynium‐22 (selective for OCT2 versus OAT2). When incubated with the RPTEC, creatinine uptake was not inhibited by indomethacin (up to 0.5 mM) and only inhibited by decynium‐22. Importantly, the decynium‐22 IC₅₀ for creatinine and metformin transport in both primary RPTEC (N = 3 donors; range 0.4 to 1.3 and 0.5 to 2.2 μM, respectively) and OCT2 transfected cell models (1.0 and 3.6 μM, respectively) was consistent. Therefore, creatinine presented as a dual (OAT2 and OCT2) substrate when incubated with transfected human embryonic kidney 293 cells, but only as an OCT2 substrate in the presence of RPTEC with functioning OAT2 activity. The authors hypothesized that the membrane potential could have impacted OCT2 activity in the transfected cells (−20 to −35 mV vs. −70 mV in RPTEC) leading to an underestimate of the uptake versus OAT2. From the standpoint of RPTEC, it is also known that OCT2 protein expression is greater than that of OAT2 (e.g., ~300 vs. ~40 pmol per gram of tissue) and possible OAT2‐OCT2 membrane heterocomplexes might impact OAT2 function [73, 74]. The above authors also concluded that if they took into account the membrane potential of the OCT2 transfected cells, and corrected for the abundance difference (transfect versus kidney tissue), that the f _t for OCT2 with creatinine would be high (~0.8) and reasonably consistent with their RPTEC data [72]. In agreement, there are some reports describing genome‐wide association studies linking OCT2 (SLC22A2) polymorphisms to serum creatinine [67, 72]. Although there are no definitive data relating creatinine renal clearance to OAT2 (SLC22A7) genotype to help validate the existing in vitro RPTEC data, it is noteworthy that most drugs known to increase serum creatinine, or decrease creatinine renal clearance, present as weak OAT2 inhibitors (versus OCT2) in vitro [69, 72]. In this regard, the OAT2 versus the OAT3 inhibition profile of indomethacin is of interest. Both transporters present similar IC₅₀s in vitro (~2 μM) but indomethacin (50 mg) decreases the OAT3‐mediated renal clearance of intravenous furosemide (~40%) without impacting serum creatinine [72]. Such clinical data argue against the role of OAT2 in creatinine clearance.

Given the broad interest in creatinine renal clearance, and the need to differentiate transporter inhibition versus renal toxicity, it is not surprising that numerous modeling exercises have been published [75, 76, 77, 78, 79, 80]. In most cases, the investigators have leveraged kinetic data obtained with singularly transfected cell lines expressing OAT2, OCT2, MATE1, and MATE2‐K and attempted to relate such data to the expression levels of each transporter in human kidney as much as possible (e.g., relative expression factor approach). Some have also considered the impact of membrane potential and bidirectional flux in the case of OCT2. In turn, the f _t for each renal transporter has been used as input into the various models along with in vitro inhibition parameters and perpetrator exposure. Although most of the authors have largely focused on MATE1 and MATE2‐K inhibition as the driver for serum creatinine increases (renal clearance decrease) following administration of agents such as trimethoprim, cimetidine, and pyrimethamine, it is notable that the f _t inputs for OAT2 and OCT2 have varied across the different publications. Some have excluded OAT2 and OCT2 completely, with a focus on MATE1 and MATE2‐K only, or have determined that OAT2 (f _t ~ 0.9) is major when compared with OCT2 (f _t ~ 0.1) [75, 76]. Others have concluded that data supporting the role of OAT2 in creatinine clearance are inconsistent and have focused on OCT2 and MATE1/2‐K only or have estimated that the contribution of OAT2 and OCT2 is similar [78, 79, 80]. Admittedly, while many drugs increase plasma creatinine by inhibiting apical MATE1 and MATE2‐K, it should be emphasized that significant inhibition of OCT2 cannot be ignored for some new molecular entities [81]. In such cases, PBPK modeling of creatinine clearance would require robust f_t input values for OCT2. As showcased by Mathialagan et al. [72], could such a value be derived independently by utilizing primary human RPTEC data? Fortunately, in vitro uptake data for other biomarkers such as N¹‐methylnicotinamide are less controversial since most reports support its selectivity for OCT2, MATE1, and MATE2‐K [80, 81, 82]. The development of robust and easy‐to‐use physiologically relevant cell‐based models of human RPTEC would enable further study on the relevance of drug transporters in the renal tubule system for small molecule drugs and biomarkers like creatinine.

4.3. Modeling the Induction of Hepatic OATP1B1/3 by Rifampicin

Rifampicin is a well‐established inducer of various drug‐metabolizing enzymes, so interest has also focused on ATP‐binding cassette (ABC) transporters and solute carriers [83, 84]. Beyond CYP3A4, the tissue biopsy data of rifampicin‐dosed subjects present induction of liver and intestinal multidrug resistance protein 2, as well as intestinal P‐glycoprotein and breast cancer resistance protein [23, 48, 83]. Such findings are consistent with conventional plated human hepatocyte, liver slice, and more recent three‐dimensional spheroid human hepatocyte data [83, 85]. Unfortunately, the induction of such ABC transporters will impact the interpretation of multi‐dose rifampicin data when it comes to DDI victims such as statins and OATP1B1/3 biomarkers. For example, rosuvastatin and atorvastatin are known breast cancer resistance protein substrates, while plasma biomarkers (e.g., coproporphyrin I and glycochenodeoxycholic acid 3‐O‐sulfate) are multidrug resistance protein 2 substrates [83, 86, 87]. Therefore, evoking the induction of liver OATP1B1/3 by rifampicin can be problematic, because in vitro data, in vivo animal data (e.g., nonhuman primate), human liquid biopsy (plasma‐derived liver sEV) and tissue biopsy data indicate no (or very weak) induction of either OATP1B1 or OATP1B3 by rifampicin when CYP3A4 is induced [26, 83]. Rifampicin auto‐induction over 2 weeks at 600 mg once daily, leading to reduced inhibition of OATP1B1/3, further complicates data interpretation [83]. This was evident in a recent letermovir DDI study report, where the authors tried to present a case for OATP1B1/3 induction by rifampicin [88]. It is notable that as well as being an OATP1B1/3 substrate, letermovir is also a substrate of rifampicin‐inducible UGT1A1 and P‐glycoprotein [23, 83]. Therefore, the described DDI between rifampicin and letermovir could be the net effect of UGT1A1 and P‐glycoprotein induction by rifampicin, the inhibition effect of rifampicin on OATP1B1/3, and letermovir itself inhibiting OATP1B1/3.

In terms of multi‐dose rifampicin as a DDI perpetrator, different groups have described models with various induction inputs for OATP1B1 and OATP1B3, and all have claimed to be successful [89, 90, 91, 92]. For example, Turk et al. reported an E _max of 0.38 for both OATP1B1 and OATP1B3, compared to an E _max of 2.5 and 9.0 for P‐glycoprotein and CYP3A4, respectively [91]. Tod et al. only considered OATP1B1 induction (fold‐change model input = 3.0) compared to no change for OATP1B3 and a fold‐change of 8.8 for liver CYP3A4 [92]. Asaumi did not differentiate OATP1B1 from OATP1B3 and lumped both transporters together (E _max = 2.3) versus a liver CYP3A4 induction E _max of 4.6 in their model [90]. More recently, Hartauer et al. described a PBPK model with a fold‐induction input of 2.3 for each of the two OATP1Bs (versus a fold‐increase of 16 for CYP3A4) [89]. As described previously, the model inputs described above are difficult to reconcile when one considers that available liquid (plasma‐derived liver sEV) biopsy and liver tissue biopsy expression data indicate minimal changes in OATP1B1 and OATP1B3 expression when CYP3A4 is induced (~4‐fold) in the same subjects following administration of rifampicin [26, 83]. Having said that, it is possible that some agents may modulate the post‐translation processing of OATP1B1 and OATP1B3 and give rise to activation of activity with an impact on drug PK while not impacting expression. It is envisioned that proteomic OATP1B1/3 analysis of plasma‐derived liver sEV will be essential if such DDI mechanisms are to be explored.

It is important to note that although the preponderance of in vitro, in vivo animal model, and human biopsy data point to minimal induction of OATP1B1/3 (versus CYP3A4) by rifampicin, clinically relevant OATP1B1/3 induction by other agents (e.g., carbamazepine and farnesoid X receptor agonists) cannot be ruled out [83]. Such considerations will be important if rifampicin is to be replaced as a routine clinical pharmacology index inducer of CYP3A4. PBPK models for such alternative inducers would need to be adapted to accommodate different induction signatures for CYP3A4, UGTs, breast cancer resistance protein, P‐glycoprotein, multidrug resistance protein 2, and OATP1B1/3.

4.4. Modeling the Inhibition of CYP1A2 by Oral Contraceptives Containing 17α‐Ethinyl Estradiol

When it comes to victim and perpetrator DDI, oral contraceptives have also received considerable attention. Of particular interest has been the differential impact of 17α‐ethinyl estradiol containing oral contraceptives on different CYP activities. Various studies with probe drugs have consistently shown that CYP1A2 and CYP2C19 are more inhibited in vivo than other CYPs such as CYP2D6, CYP3A4, CYP2C9, and CYP2C8 [93]. To date, however, it has not been possible to rationalize the results based on available in vitro data [94, 95, 96]. 17α‐Ethinyl estradiol is a low dose drug (~30 μg) and its plasma levels are very low compared to reported in vitro CYP IC₅₀ (K _i) values [93]. Importantly, there is no evidence for time‐dependent (mechanism‐based) inhibition of CYP1A2 and CYP2C19 in vitro, and down‐regulation of their expression has been hypothesized [93, 95].

Lack of definitive in vitro mechanism‐based inhibition data has not deterred investigators from developing PBPK models that have claimed to successfully describe clinical DDI with orally dosed CYP1A2 inhibition victims like caffeine and tizanidine [97]. Both drugs are well absorbed (~100%) and present as CYP1A2 substrates (f _m ~ 0.95), although the oral bioavailability of tizanidine is lower (21% versus 98%) because of greater intestinal (61% vs. < 5%) and hepatic (46% vs. 6%) first pass metabolism [93]. Unfortunately, the described DDI predictions have necessitated the use of unrealistic k _inact input values that are ~3 to 4 orders of magnitude greater than those measured in vitro for clinically relevant and definitive CYP1A2 mechanism‐based inhibitors (100 to 200 min⁻¹ vs. 0.01 to 0.9 min⁻¹, Table S5) [97]. Such model parameterization is difficult to rationalize and potentially suggests mis‐specification of the model structure or the use of implausible parameters that can compromise credibility assessment and predictivity of the model.

As shown in Figure 5, it is possible to empirically model the inhibition of CYP1A2. For agents such as caffeine and theophylline, their oral bioavailability is high and their extraction in the gut is negligible. Therefore, one can relate the CYP1A2 fraction activity remaining in the liver (term B in Figure 5) to the victim drug AUC ratio. For drugs like theophylline and caffeine, AUC ratios in the range of 1.5 to 2 have been reported [93]. To generate such ratios, one would need to decrease hepatic CYP1A2 activity by ~50% (term B~0.5 Figure 5). As has been hypothesized, such a decrease in activity could result from the attenuation of liver CYP1A2 expression involving the estrogen receptor. At clinically relevant doses, 17α‐ethinyl estradiol is a potent estrogen receptor agonist which could lead to repression of CYP1A2 and CYP2C19, as well as induction of enzymes such as UGT1A4 [93]. The availability of tissue or liquid biopsy expression profiling data (protein, mRNA, and miR) from an oral contraceptive DDI study would enable the testing of such a hypothesis. If available, such data could then be used to validate PBPK model‐based changes in CYP1A2, CYP2C19, and UGT1A4 activity upon multiple dose administration of 17α‐ethinyl estradiol containing oral contraceptives.

FIGURE 5.

Inhibition of CYP1A2 and its impact on victim drug AUC ratio. (A) Reported plasma AUC ratios for oral caffeine and theophylline when co‐administered with 17α‐ethinyl estradiol‐containing oral contraceptives [93]. (B) Relating the AUC ratio to the fraction of liver CYP1A2 activity remaining. The described model tests two values for CYP1A2 f _m (0.85 and 0.95) and was adapted from the model described in Figure 2 (intestine is ignored and induction not considered for 17α‐ethinyl estradiol‐containing oral contraceptives). The shaded area represents the reported AUC ratio range and the corresponding fraction activity remaining [93]. CYP1A2, cytochrome P450 1A2; f _m, fraction metabolized by CYP1A2; B, fraction activity remaining in the liver (fraction inhibited = 1 − B, Figure 2); AUC, area under the plasma concentration versus time curve; AUC ratio, area under the plasma concentration versus time curve (AUC) ratio (AUCR = AUC_{oral contraceptive}/AUC_reference).

4.5. Modeling of Pregnancy‐Related Changes in Liver CYP2D6 Expression and Function

Pregnancy is known to impact the activity of various drug‐metabolizing enzymes and this is reflected in numerous manuscripts that have reported trait measures such as probe drug AUC, metabolite‐to‐parent drug ratio (urine and/or plasma), and biomarkers like plasma 4β‐hydroxycholesterol (Table S6 reference 1) [98, 99]. Attempts have also been made to evaluate transporter function using drug probe substrates (e.g., P‐glycoprotein and OATP1B1/3, Table S6 reference 2 and 3). Such clinical data have spawned various efforts to model CYP3A4, CYP1A2, and CYP2D6 activity during pregnancy and post‐partum (Table S6 references 4 to 7). In most cases, such models have described decreased CYP1A2 activity, accompanied by the steady time‐dependent increase in CYP3A4 and CYP2D6 activity throughout pregnancy (versus post‐partum or non‐pregnant females). For example, default settings in commercially available PBPK software present a 1.36‐ (1.25‐), 2.16‐ (1.75‐), and 3.1 (2.2)‐fold increase in CYP2D6 (CYP3A4) function at the end of the first, second, and third trimester, respectively (Table S6 reference 5).

Unfortunately, there are minimal longitudinal studies with individual pregnant subjects, and no reports are available describing biopsy data. To date, there is only one report describing the proteomic analysis of serum‐derived liver sEV of a small number of pregnant subjects (N = 3 subjects; only trimester information available not the specific month of pregnancy) [26]. In this instance, the authors measured various pregnancy hormones in the serum samples used to isolate the liver sEV to ensure the pregnancy trimester phenotype of the individuals (Figure S4) and showed that pregnancy did not impact liver sEV OATP1B1 or OATP1B3 protein expression and that CYP2D6 protein expression was not significantly increased in the first and second trimester (1.2‐ and 1.0‐fold, respectively, Figure S4). Only in the third trimester was CYP2D6 expression significantly increased (3.7‐fold, p = 0.03) versus non‐pregnant subjects. For CYP3A4, protein expression did trend higher throughout pregnancy (1.8‐ and 1.9‐fold increase in the first and second trimester, respectively) but only in the third trimester samples did the increase (3.2‐fold) reach statistical significance (p = 0.003) versus non‐pregnant females (Figure S4). Of note, CYP3A4 protein expression in the serum‐derived liver sEV did relate to the concentration of both progesterone and estrogen in the serum of the same subjects [26]. The increase in CYP2D6 expression in the third trimester (versus first and second trimester) is not unexpected because others have reported a similar profile in pregnant humanized CYP2D6 mice (gestation day 21 vs. days 7 and 14; Table S6 reference 8) [26]. In this regard, CYP2D6 differs from CYP3A4 because of different transcription factors regulating its expression (Table S6 references 1 and 9).

After consideration of the above, it is concluded that more information is needed regarding the temporal changes in CYP2D6 expression and function during pregnancy. Such data would inform PBPK model building and circumvent the use of CYP2D6 drug probe‐based trait measures that may be affected by changes in CYP3A4 activity. Ideally, non‐invasive methods such as a CYP2D6 biomarker (e.g., 3,4‐seco‐solanidine‐3,4‐dioic acid), or measurement of CYP2D6 expression in plasma or serum‐derived sEV, would be deployed in a formal longitudinal study with readouts every few days or every week [26, 36]. Such a concept is exemplified by O'Brien et al., who reported the time‐dependent measurement of the venlafaxine metabolite ratio in hair samples of a single pregnant subject (Figure S4, Table S6 reference 10). Venlafaxine is known to be extensively metabolized by CYP2D6, with f_m estimated to be 0.77 to 0.83 based on reported poor metabolizer (PM) versus extensive metabolizer (EM) plasma AUC ratio [f _m = 1‐(1/AUC ratio); AUC ratio = AUC_PM/AUC_EM = 4.3 to 5.9; Table S6 references 11 and 12]. The data reported by O'Brien et al., although limited, are interesting because they showed that the venlafaxine metabolite ratio was increased in the third trimester and decreased post‐partum as expected. Such results are consistent with published PBPK models (Table S6 references 4 to 7). However, the metabolite ratio only really increased in the last month of the second trimester (Figure S3), which is not reflected in currently published PBPK models. What is widely described in published PBPK models is the ~2‐ to 3‐fold increase in both CYP3A4 and CYP2D6 activity in the third trimester consistent with available plasma‐derived liver sEV proteomic data (Table S6 references 4 to 7) [26]. Beyond the CYPs, modeling of liver OATP1B1/3 changes in pregnancy may also be challenging. Although some have reported decreased activity, sEV‐based liquid biopsy data do not support any statistically significant change in maternal liver OATP1B1/3 protein expression throughout pregnancy (Table S6 references 2 and 3) [26]. Unfortunately, there is also a paucity of information related to other drug‐metabolizing enzymes in the liver (e.g., UGTs, sulfotransferases) and there are no data for extrahepatic organs such as the intestine. Therefore, investigators may turn to the liquid biopsy approach (e.g., broad proteomic profiling of plasma or serum‐derived liver and intestinal sEV) to fill in these gaps in knowledge and support pregnancy‐PBPK model building. Such efforts will not only help elevate our understanding of DDI in pregnant patients but also support dosing recommendations for experimental drugs with validated PBPK models.

5. Concept of Using PBPK Model‐Independent Data to Support DDI Model Verification

DDI are complex, and many parameters can jointly govern clinical outcomes (Figure S1). Such complexity is what makes PBPK modeling of DDI so powerful, but it comes with challenges, especially if there is a degree of uncertainty or inconsistency when utilizing various input data. This is very evident for in vitro “drug‐dependent inputs” (e.g., subcellular fraction‐ and cell‐based assay‐derived K _i, k _inact, E _max, and EC₅₀), which often do not translate directly to in vivo and require the application of scaling factors, use of calibrators, and correction for unbound fraction in vitro. In addition, predicting the exposure of the perpetrator (e.g., free concentration versus time profile) in target tissues is not trivial, especially when having to consider its intracellular free fraction and the influence of any active transport in vivo (free concentration in tissue ≠ free concentration in plasma). Then there are challenging “system inputs” describing various physio‐ and pathophysiologic endpoints such as the enzyme (e.g., CYP3A4) degradation rate constants in the gut and liver. Therefore, even if the plasma PK of the perpetrator is modeled successfully, a good DDI prediction is not guaranteed. As described in the present narrative (Example 4), there may also be instances when a PBPK‐based DDI prediction is “good” but necessitated the use of seemingly implausible model parameters that are inconsistent with values widely published in the literature. In such cases, it is possible that while the model may describe the data that was used to build it, it could lose productivity when challenged with a new and unstudied scenario, like a different victim drug. Such challenges will only be exacerbated if attempts are made to use the same models to predict DDI in drug‐metabolizing enzyme or transporter genotyped or target subjects (e.g., pregnant, diseased, geriatric, obese, pediatric, organ impaired), and when trying to incorporate inhibitory metabolites or complex mechanisms (inhibition plus induction).

Given the above challenges, it is not surprising that PBPK modeling consensus guidelines and best practices have been authored and various investigators have called for improved PBPK model and parameter verification and the establishment of credibility assessment frameworks [16, 17, 18, 19, 20, 100]. Such publications have described PBPK model uncertainty, establishing confidence, software qualification, model refinement, model verification, sensitivity analysis, and assessment of prediction performance, all with a focus on model parameters (system‐ and drug‐dependent) and assumptions with data sources, as well as an acknowledgment that some parameters need to be derived and fitted to support the modeling exercise. One possibility is to leverage relevant data that have been generated independent of the PBPK modeling exercise to support the assignment of parameter values. The concept has already been illustrated by the reports of Yamazaki et al. [48] and Quinney et al. [49], who leveraged tissue biopsy data to validate their PBPK models. As described for CYP3A4 with inducers and CYP2D6 in pregnancy, could tissue and liquid biopsy data also be used to verify PBPK model inputs for induction DDI models? Likewise, as illustrated by creatinine, independently generated in vitro data from validated primary human RPTEC should be used, along with transfected cell data, to estimate the role (f_t) of different solute carriers in renal clearance (versus glomerular filtration). This could be extended beyond primary human hepatocytes to also include primary human enterocytes to support the generation of f_m and f_t signatures of various ABC, solute carrier, and drug‐metabolizing enzyme substrates.

6. Conclusions

In the era of precision medicine, polypharmacy, and ever‐increasing computational power, considerable effort has been made to develop static and dynamic DDI prediction models with the aim of using the output of validated models for dosing recommendations in drug labels in lieu of conducting formal clinical DDI studies [8, 9, 10, 11]. Over time, this has encompassed more complex quantitative PBPK models that not only predict DDI, but also gene‐DDI, and even disease‐DDI and beyond. Such progress has been made possible by continuous research in both the academic and private sector and the greater availability of various software packages enabling the use of complex PBPK models that have increasingly supported sponsor‐initiated submissions to agencies, reflected in recently issued ICH documentation (ICH M12) and the growing number of peer‐reviewed publications describing PBPK‐based DDI predictions [4, 5, 12, 13, 15, 16]. Although considerable progress has been made, the various challenges and pitfalls of PBPK modeling have been described, and there has been a call for greater confidence in the models, improved model verification, and credibility assessment [7, 17, 18, 19, 20, 21, 22].

As reviewed herein, the need for model verification is apparent when using in vitro data to model induction and reporting successful DDI predictions. For example, evoking induction by rifampicin involving greater than 8‐fold increases in liver or intestinal CYP3A4 function, as well as PBPK model E_max input values of 2 to 3 for liver OATP1B1/3, should be questioned [52, 62, 63, 64, 65, 66, 89, 90, 92]. Such considerations are important as sponsors are now having to switch to alternative inducers like carbamazepine. But even for carbamazepine, published PBPK models are discordant regarding the fold increase in gut CYP3A4, although available biopsy data present a ~2.5‐fold increase in gut and liver CYP3A4 expression that can support non‐PBPK predictions of oral midazolam AUC ratio and plasma biomarker changes (Figures 3 and 4, Figure S3) [59, 61].

Beyond induction, the additional examples described in the present narrative further illustrate the challenges faced by investigators when attempting to model complex DDI using in vitro data. Here also, in the absence of more robust mechanistic information, some of the author's model inputs could be questioned. This was apparent when trying to model DDI with 17α‐ethinyl estradiol and ascribing the clinically observed decrease in CYP1A2 activity to mechanism‐based inhibition with a very high k _inact (~100 min⁻¹) or assigning a high input f_t for OAT2 (~0.9) when predicting inhibition of creatinine renal clearance [76, 97]. The same also applies to the modeling of pregnancy‐related changes in CYP2D6 function, especially during the second trimester, which will remain challenging in the absence of robust longitudinal data obtained with a selective CYP2D6 biomarker and profiled plasma‐derived liver sEV samples (Table S6 references 4 to 7) [26, 36].

Without a doubt, the last decade has seen the rapid development and implementation of biomarkers and isoform selective probe substrate strategies (e.g., cocktails and microdoses) facilitating the generation of clinical DDI study packages to enable decision making, sponsor‐initiated submissions, and informed dosing recommendations in drug product labels [24]. More recently, this has extended to the exploration of plasma‐derived sEV as liquid biopsy and increased use of primary human tissue cell culture models (hepatocytes, RPTEC, enterocytes) to enable the generation of in vitro parameters under more physiologically relevant conditions (e.g., microphysiological systems, MPS, and three dimensional tissue culture) [23, 24, 25, 26, 27, 28, 29, 30, 72, 85]. Data from such in vitro models could be modeled with adapted versions of PBPK software (Table S7). Therefore, with further development and validation, it is envisioned that liquid biopsy, clinical biomarker, and MPS‐PBPK could be used in a supportive and complementary fashion to validate both drug and system parameters used in static and PBPK‐based DDI prediction models.

Concomitantly, computing power has increased dramatically so that PBPK modeling of complex victim and perpetrator DDI can be attempted [4, 5, 13, 14]. However, as computer science progresses at an even more rapid pace, it is envisioned that such DDI modeling will increasingly encompass advancements in data analysis and artificial intelligence (Table S8) [20]. Therefore, could artificial intelligence be incorporated into PBPK software to support unbiased optimization of model parameters and improve links to independent sources of literature data? [20] Could such linkages enable seamless integration of model‐independent data sources, such as those already described, with PBPK model‐derived output and input parameters to strengthen model verification and build model credibility? Such efforts are worthwhile because of the inevitable acceleration in computational power leading to the development of even more complex DDI models. In turn, such complex DDI models will necessitate even more rigorous validation and credibility assessment approaches. Beyond technology, and equally as important, is the continued growth in knowledge related to different DDI mechanisms. As illustrated by the rifampicin‐zidovudine DDI and 17α‐ethinyl estriol‐dependent inhibition of CYP1A2 [50, 93], PBPK modeling is only useful if one fully understands the fundamental mechanisms underpinning the observed interaction.

Disclosure

David Rodrigues and Christopher Gibson are employees and stockholders of Incyte Corporation, Wilmington, DE, USA. No artificial intelligence (AI) or AI‐related tools were used in the preparation of this manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1.

Rodrigues D., Gibson C. R., and Isoherranen N., “Drug Interaction PBPK Modeling: Review of the Literature Exposes the Need for Increased Verification of Model Inputs and Outputs as Part of Credibility Assessment,” Clinical and Translational Science 18, no. 7 (2025): e70299, 10.1111/cts.70299.