Building Confidence in Physiologically Based Pharmacokinetic Modeling of CYP3A Induction Mediated by Rifampin: An Industry Perspective

Abstract

Physiologically‐based pharmacokinetic (PBPK) modeling offers a viable approach to predict induction drug–drug interactions (DDIs) with the potential to streamline or reduce clinical trial burden if predictions can be made with sufficient confidence. In the current work, the ability to predict the effect of rifampin, a well‐characterized strong CYP3A4 inducer, on 20 CYP3A probes with publicly available PBPK models (often developed using a workflow with optimization following a strong inhibitor DDI study to gain confidence in fraction metabolized by CYP3A4, f _m,CYP3A4, and fraction available after intestinal metabolism, Fg), was assessed. Substrates with a range of f _m,CYP3A4 (0.086–1.0), Fg (0.11–1.0) and hepatic availability (0.09–0.96) were included. Predictions were most often accurate for compounds that are not P‐gp substrates or that are P‐gp substrates but that have high permeability. Case studies for three challenging DDI predictions (i.e., for eliglustat, tofacitinib, and ribociclib) are presented. Along with parameter sensitivity analysis to understand key parameters impacting DDI simulations, alternative model structures should be considered, for example, a mechanistic absorption model instead of a first‐order absorption model might be more appropriate for a P‐gp substrate with low permeability. Any mechanisms pertinent to the CYP3A substrate that rifampin might impact (e.g., induction of other enzymes or P‐gp) should be considered for inclusion in the model. PBPK modeling was shown to be an effective tool to predict induction DDIs with rifampin for CYP3A substrates with limited mechanistic complications, increasing confidence in the rifampin model. While this analysis focused on rifampin, the learnings may apply to other inducers.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

PBPK modeling is increasingly used for the assessment/prediction of DDIs. For model‐strong CYP3A4 inducer rifampin, clinical DDIs with multiple CYP3A substrates have been reported. PBPK modeling has been conducted to evaluate rifampin‐CYP3A substrate DDIs.

• WHAT QUESTION DID THIS STUDY ADDRESS?

This work utilized a wealth of high‐quality published data to assess DDI simulation performance for predicting the effect of CYP3A inducer rifampin on CYP3A4 substrates with varied features. The learnings should apply to other inducers (e.g., other CYPs, moderate inducers).

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

Given a sufficient data package and the use of an appropriate workflow, PBPK modeling may be used to accurately predict the effect of rifampin on certain types of well‐characterized CYP3A substrates. Scenarios are identified in which PBPK modeling may accurately define the magnitude of such effects, potentially in lieu of a clinical study.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

Strengthening our understanding of the confidence to predict rifampin DDIs with CYP3A substrates potentially enables a shift in clinical study planning such that a clinical DDI study could be performed with a moderate instead of a strong inducer, followed by PBPK modeling of the strong inducer effect. A moderate inducer study may be preferred due to its clinical relevance or the need to conduct a study in a patient population. This option is particularly important considering the inability to use rifampin in healthy volunteers due to nitrosamines.

Physiologically‐based pharmacokinetic (PBPK) modeling offers a viable approach to predict changes in exposure of a new chemical entity due to drug–drug interactions (DDIs) with the potential to streamline or reduce clinical trial burden. PBPK models are part of model‐informed drug development (MIDD), and establishing model credibility is critical for regulatory decision‐making. ¹ Both the European Medicines Agency (EMA) and the United States Food and Drug Administration (FDA) have issued guidance outlining requirements for reporting PBPK models including methods for demonstrating the scientific basis/strength of the modeling. ² , ³ , ⁴

The International Consortium for Innovation and Quality in Pharmaceutical Development (IQ) established a working group under the Translational ADME Leadership Group (TALG) with the aim to build confidence in using PBPK modeling to predict cytochrome P450 (CYP) 3A4 induction DDIs. Previously, the IQ‐PBPK modeling induction working group (PBPK‐IWG) collected case studies and conducted a survey to understand general strategies and regulatory submission experience for PBPK modeling of CYP3A induction. ⁵ This previous work highlighted areas of opportunity and a workstream was developed to apply a platform‐agnostic approach to evaluate the ability of PBPK modeling to predict clinical outcomes for CYP3A substrates with varying attributes (i.e., fraction metabolized by CYPs [f _m], fraction available after intestinal metabolism [Fg], substrate of P‐glycoprotein [P‐gp], substrate of other inducible pathways, and whether compounds are susceptible to autoinduction or autoinhibition or impact the perpetrator PK). Rifampin, a well‐studied, prototypical strong CYP3A4 inducer, was selected as the reference inducer because of its historical widespread use in clinical DDI studies, extensive in vitro data on its DDI properties, and availability of substrate PBPK models.

Rifampin has a complex DDI profile. It induces multiple enzymes and transporters in addition to CYP3A4 and CYP3A5, including CYP1A2, 2B6, 2C8, 2C9, 2C19, UDP‐glucuronyltransferases (UGTs), sulfotransferases, carboxylesterases, and transporters including P‐gp and multidrug resistance‐associated protein 2 (MRP2). ⁶ , ⁷ Rifampin also inhibits several enzymes and transporters, including CYP3A and 2C8, breast cancer resistance protein (BCRP), P‐gp, Na‐dependent taurocholate co‐transporter (NTCP), MRP2, and the organic anion transporting polypeptide isoforms OATP1B1 and 1B3. ⁸ , ⁹ , ¹⁰ , ¹¹ Coadministration with an OATP substrate may mask the magnitude of clinical induction as has been observed with repaglinide. ¹² Rifampin use is currently precluded in healthy volunteers due to higher than acceptable limits of nitrosamines. Hence building confidence in PBPK modeling to predict DDIs and inform drug labeling of CYP3A substrates offers an alternative approach for drug development. ⁵ , ¹³ , ¹⁴ This strategy was recently used successfully for vonoprazan, where instead of conducting a strong inducer clinical DDI study, PBPK simulations were used to inform drug labeling. ¹⁵ , ¹⁶

PBPK models for rifampin are available in several software platforms (e.g., Simcyp, PK‐Sim, and GastroPlus). ¹⁷ , ¹⁸ , ¹⁹ For the Simcyp model, iterative optimizations have been made to improve the known underprediction of induction magnitude. ¹⁷ Factors implicated in the underprediction include variability of in vitro and in vivo data, limited validation of substrate models, rifampin pleiotropic effects, differing in vitro–in vivo extrapolation (IVIVE) scaling approaches, and clinical study design. ¹⁷ , ²⁰ In addition to modeling the pleiotropic effects of rifampin, appropriate parameterization of substrate disposition pathways is critical for the development of high‐fidelity PBPK models to predict induction DDIs. Both a good substrate model and a good precipitant model are needed to deliver mechanistic insights via PBPK modeling. In several literature examples, the rifampin PBPK model used for initial predictions failed to recover the observed magnitude of clinical induction, and additional attributes of the substrate drug needed to be incorporated. For example, in recent models for P‐gp substrates, the effect of P‐gp on permeability and extent of absorption and the impact of induction and/or inhibition of P‐gp at the gut, liver, and/or kidney have been incorporated and have improved the accuracy of the PBPK model. ²¹ , ²²

Complex disposition of an object substrate can affect PBPK simulation of induction potential, requiring empirical adjustment of the rifampin model. However, for some substrates with fewer mechanistic complications, PBPK modeling using a standard rifampin model may be suitably predictive of DDIs. The PBPK‐IWG attempted to understand the predictivity of PBPK modeling using rifampin base models with multiple victim CYP3A and dual CYP3A/P‐gp substrates. A primary goal of this analysis was to determine whether the workflow proposed in our recent publication would result in accurate predictions for the effect of rifampin on CYP3A probe substrates. ⁵ Here, predicted and observed DDIs are compared and recommendations are presented for future induction DDI predictions using PBPK modeling.

METHODS

Identification of CYP3A substrate models and selection criteria

Available PBPK models for CYP3A substrates were identified through literature searching, knowledge of PBPK‐IWG group members, and assessing models available in PBPK modeling software packages from 2020 to 2021. Initially, models from Simcyp™, GastroPlus®, and PK‐Sim® were considered. Simcyp yielded the majority of CYP3A substrate models, and therefore was chosen for the present work.

CYP3A substrates with a Simcyp PBPK model and clinical repeat‐dose rifampin DDI data available were assessed according to the previously published substrate workflow. ⁵ Inclusion was based on the availability of human absorption, disposition, metabolism, and excretion (ADME) study data, DDI risks for the compound both as a substrate (i.e., possible additional mechanisms by which rifampin could cause a DDI were assessed) and precipitant (i.e., additional mechanisms by which the CYP3A substrate could impact its own PK or rifampin PK was assessed). The dataset included substrates with a wide range of fractions metabolized by CYP3A4, f _m,CYP3A4, and Fg values (Table 1 ). For most drugs in the analysis, only oral (PO) clinical DDI data were available with rifampin. However, nifedipine, midazolam, and alfentanil had data available for both intravenous (IV) and PO administration with rifampin.

Table 1.

Key physiochemical and PK parameters for CYP3A substrate PBPK models ^a

Drug	MW	LogP	Single agent PBPK modelg					With induction by rifampin
			f m,CYP3A4	F	Fa	Fg	Fh	f m,CYP3A4	F	Fa	Fg	Fh	P‐gp, appb	Mechanistic considerations
Tier 1 Substrates
First‐order absorption models
Abemaciclib	506.6	3.36	0.89	0.45	0.91	0.73g	0.68	0.97	0.022	0.91	0.16	0.15	✓, M	P‐gp inhibitor
Alfentanil	416.5	2.16	0.92	0.33	0.99	0.52 (0.54c)	0.64	0.99	0.028	0.99	0.19	0.15	‐, H
Alprazolam	308.8	2.12	0.69	0.83	0.87	0.995 (0.86c)	0.96	0.86	0.59	0.87	0.97	0.70	‐, H
Bosutinib	530	3.1	0.98	0.25	0.66	0.49	0.76	0.99	0.016	0.66	0.098	0.25	✓, M	A later model included P‐gp in ADAM model
Doravirine	425.7	3.0	0.82	0.60	0.63	0.99	0.96	0.97	0.42	0.63	0.95	0.70	✓, H
Eliglustat	404.5	2.84	0.086	0.06	0.94	0.55	0.11	0.35	0.010	0.94	0.20	0.055	✓, H	f m,CYP2D6 0.81, fe 0.07, CYP2D6 auto‐inhibitor
Eliglustat in CYP2D6 PMs	404.5	2.84	0.86	0.29	0.94	0.55	0.55	0.91	0.019	0.94	0.20	0.10	✓, H	f m,CYP2D6 0, fe 0.14
Midazolamf	325.8	3.53	0.81d	0.30	0.99	0.57 (0.57c)	0.53	0.82	0.019	0.99	0.20	0.10	‐, H
Naloxegol	652	1.43	0.90	0.28	0.66	0.87	0.49	0.94	0.03	0.66	0.58	0.08	✓, M	Empirical approach for P‐gp
Nifedipine	346.3	2.69	0.94	0.41	0.99	0.65 (0.68c)	0.63	0.94	0.036	0.99	0.27	0.14	‐, H	P‐gp inh., moderate (not strong) CYP3A inh. diltiazem DDI study available
Triazolam	343.2	2.03	0.85d	0.50	0.87	0.73 (0.67c)	0.79	0.86	0.075	0.87	0.33	0.26	‐, H
ADAM models
Ibrutinib	440.5	3.97	0.95	0.032	0.89	0.44 (0.47c)	0.088	0.99	0.0012	0.89	0.16	0.0087	✓, H
Tier 2 Substrates
First‐order absorption models
Aprepitant	534.4	4.80	0.86	0.39	0.87	0.48	0.93	0.97	0.075	0.87	0.16	0.54	✓, H	Autoinduction, auto‐TDI
Ethinyl estradiol	296.4	3.81	0.20	0.41	0.95	0.50	0.87	0.52	0.20	0.95	0.35	0.60	✓, H	Not all elimination pathways characterized, non‐CYP3A pathways contributing to intestinal and hepatic metabolism, additional inducible pathways 23
Osimertinib	499.6	5.45	0.43	0.67	0.75	0.996	0.90	0.71	0.042	0.75	0.10	0.56	✓, H	CYP3A4 TDI, weak inducer
Simvastatin	418.6	4.68	0.89	0.032	0.87	0.11 (0.08c)	0.34	0.98	0.0010	0.87	0.025	0.048	‐, H	OATP substrate
Tofacitinib	312.4	1.15	0.51	0.68	0.83	0.98	0.83	0.79	0.27	0.83	0.90	0.36	✓, H	f m,CYP2C19 0.17
ADAM models
Everolimush	958.3	5.56	1.0	0.55	0.87	0.72	0.87	1.0	0.13	0.87	0.38	0.38	✓, L	P‐gp not explicitly included, BPR decreases at high concentrations but typically concentrations are low enough that it does not impact PK
Olaparib	434.5	1.55	0.72	0.91	1.0	0.995	0.92	0.93	0.51	1.0	0.99	0.51	✓, L	P‐gp inh., CYP3A4 TDI
Ribociclib	434.5	1.95	0.69	0.56	0.99	0.99	0.57	0.93	0.13	0.99	0.95	0.14	✓, M	CYP3A4 TDI
Venetoclaxd	868.4	7.9	1.0	0.58d	0.78	0.99	0.74	1.0	0.17d	0.78	0.95	0.23	✓, L	P‐gp inh., OATP inh., P‐gp not explicitly included

ADAM, advanced dissolution, absorption, and metabolism; BCRP, breast cancer resistance protein; BPR, blood‐to‐plasma ratio; CL, clearance; Cp, plasma concentration; F, bioavailability; Fa, fraction absorbed; fe, fraction excreted in urine; Fg, fraction that escapes intestinal first‐pass metabolism; Fh, fraction that escapes hepatic first‐pass metabolism; f _m,CYPX, fraction of systemic clearance through a given CYP enzyme based on total clearance; H, high; inh., inhibitor; L, low; M, moderate; OATP, organic anion transporting polypeptide; P‐gp, P‐glycoprotein; TDI, time‐dependent inhibitor.

^a For simulation results, the geometric mean for the 300–500 virtual subjects in the trial was reported. For first‐order absorption model simulations with rifampin coadministration, F and Fa were not included in simulation results, and for ADAM model simulations F was not included in simulation results. The rifampin simulation individual parameter values were calculated as: F = CL/CLpo, and (for first‐order absorption models) Fa = F/Fh/Fg. For compounds that had IV and PO DDI studies, the PO PK models were used here.

^b The assessment of whether a substrate is a substrate of P‐gp was based on in vitro data, for example, based on the drug label or a published study. Additional details are found in Table S1 .

^c The number in () is an observed Fg. For ibrutinib, the observed Fg is calculated from a reported grapefruit study. ²⁴ For all others, the observed Fg is the value used for Qgut model development as described by Yang et al. ²⁵ For ethinyl estradiol, a grapefruit juice study was available, but since ethinyl estradiol is a substrate of several enzymes expressed in enterocytes including CYP3A4, CYP2C9, UGT1A1, and SULT1E1, the grapefruit juice study data cannot be used to determine Fg in a typical manner (i.e., based on the assumption that grapefruit juice inhibits all the CYP3A in the intestines completely with a minimal effect on CYP3A in the liver and that it is primarily intestinal metabolism by CYP3A leading to Fg less than 1). ²³ , ²⁶

^d For midazolam and triazolam, the geometric mean values for f _m,CYP3A4 are misleading. The median values are 0.96 and 0.97, respectively. The difference is due to the specific incorporation of CYP3A5 in the model. Other models specifically incorporated CYP3A5 (e.g., nifedipine and osimertinib), but there is not such a difference in the f _m,CYP3A4.

^e Calculated using geometric mean values as Fa × Fg × Fh.

^f Many midazolam studies were available. The simulations here were for the study of Adams et al., with a 2 mg midazolam dose. ²⁷

^g For abemaciclib, the Fg value is for the 200 mg dose; for 50 mg dose, Fg was set to a value of 1.

^h For everolimus, values are presented using an E _max = 16 for rifampin; however, model predictions were more aligned with clinical observations with E _max = 8 due to its development in an earlier version of Simcyp.

Certain mechanisms/PK properties/DDI properties were expected to confound the modeling of DDIs with rifampin and so substrates were binned into two tiers according to these properties. Tier 1 included compounds that were more selective CYP3A substrates and were not inhibitors or inducers of CYP3A. OATP substrates were not included in Tier 1. Tier 2 compounds were associated with greater complexity (e.g., OATP substrates and inhibitors and inducers of CYP3A were allowed). P‐gp substrates were included in both Tier 1 and Tier 2. Table 1 lists the reasons CYP3A substrates were considered Tier 2 in the “Mechanistic considerations” column.

CYP3A substrate models with an insufficient data package, significant gaps in mechanistic understanding incorporated in the substrate model, or high uncertainty or complexity were excluded from the analysis. For example, the analysis did not include the CYP3A substrate tadalafil due to limited data available for PBPK model development and did not include zolpidem due to its relatively large component of metabolism by CYP2C9 where there is lower confidence in induction extrapolation.

PBPK models for CYP3A substrates

PBPK models are a type of PK model that represents the body as well‐stirred tanks for organs connected by blood flow with model parameters based on physiological properties, such as volumes, blood flow rates, and tissue compositions. ²⁸ Even before the first‐in‐human clinical study, these models can be parameterized based on data typically generated during nonclinical drug development and used to predict human PK. ²⁹ Once human PK data are available, the confidence in the model is improved and the model can be applied for a range of useful applications, for example, DDI predictions. ³⁰

This analysis uses PBPK models for CYP3A substrates already available in the literature, the Simcyp software, and other publicly available sources. The sources and key details for these models are provided in Table S1 , with a focus on identifying any data gaps, limitations or features of interest. Any changes that needed to be made to the model (e.g., in the case where a model was developed in a much earlier version of the software and no longer resulted in good agreement with the current version) were documented.

Most substrate PBPK models in this analysis applied a first‐order absorption model, but five substrate models used the mechanistic advanced dissolution and metabolism (ADAM) model (Table 1 ). The first‐order absorption model within Simcyp incorporates intestinal metabolism using the Qgut model. The Qgut model is a semi‐empirical mathematical model incorporating permeability through the enterocyte membrane and villous blood flow to describe first‐pass intestinal metabolism. ³¹ The equations for the Qgut model shown here were used for the CYP3A substrate models with first‐order absorption. Eqs. 1 and 2 describe the Qgut model and how Fg is impacted when this model ³¹ is employed:

$$Q_{\mathrm{gut}}=\frac{Q_{\text{villi}}*{\mathrm{CL}}_{\text{perm}}}{Q_{\text{villi}}+{\mathrm{CL}}_{\text{perm}}}$$ (1)

$$F_{\mathrm{g}}=\frac{Q_{\mathrm{gut}}}{Q_{\mathrm{gut}}+{\mathrm{fu}}_{\mathrm{G}}*{\mathrm{CL}}_{\mathrm{u},\mathrm{int},\mathrm{G}}}$$ (2)

where Q _villi is villous blood flow, fu_G is the fraction unbound in the gut wall (i.e., in enterocytes), CL_u,int,G is the unbound intrinsic metabolic clearance in the gut, and CL_perm is a clearance term for permeability through enterocytes. Substrate PBPK models using the ADAM model (with separate intestinal compartments for duodenum, jejunum, ileum, and colon) ³² incorporate a more mechanistic description of intestinal permeability and first‐pass metabolism.

After oral administration of a compound, the luminal concentration is relatively high, with potential DDI implications. Based on the absorption characteristics of both compounds, their concentration could be high around the same time or at different times which would affect the extent of competitive inhibition. The selection of first‐order or ADAM absorption models in the PBPK model for the substrate or inducer influences the pertinent interacting concentrations and impacts the prediction of intestinal metabolism and DDIs such as induction by rifampin. The time‐dynamics of luminal concentration of both substrate and inducer are a feature of these PBPK models.

When either substrate or inducer or both have first‐order absorption models, the relevant interacting concentration for induction in the intestines (i.e., the estimated concentration of inhibitor or inducer in the enterocyte, [I]_g) is the unbound drug concentration in the hepatic portal vein, given by fu_G*Cpv, where Cpv is the total drug concentration in the portal vein, which can also be calculated as Fa*ka*Dose/Q _ent, where ka is the first‐order absorption rate constant and Q _ent is the estimated enterocyte blood flow. ³³ Since the rifampin PBPK model used in this analysis incorporates a first‐order absorption model, the interacting drug concentration in the intestine is always estimated as fu_G*Cpv, i.e., a surrogate of drug concentration in enterocytes, irrespective of the selection of absorption model for the substrate. For substrate models employing an ADAM absorption model, regional absorption and regional CYP3A4 abundance are considered for the PK of the substrate; however, the interaction is simulated to occur from the unbound portal vein concentrations of rifampin and the resulting CYP3A induction's effect on substrate metabolism.

This analysis aimed to assess whether using the previously proposed workflow ⁵ would result in accurate DDI predictions for the effect of rifampin on CYP3A substrates. For some earlier models, there tends to be more clinical data available for model development, for example, a grapefruit juice study useful for model parameterization of Fg, which provides different (potentially more informative) data for model development than is required by the proposed workflow. Also, for models developed iteratively over the years, the process followed for model development is not always clear. Some of these CYP3A substrates were used for both Qgut model development and rifampin model optimization (i.e., alfentanil, alprazolam, midazolam, nifedipine, simvastatin, and triazolam). Therefore, a subset of models that were not used in Qgut or rifampin model development used the first‐order absorption and Qgut model, and had documentation that the proposed workflow was generally followed, were identified (i.e., abemaciclib, aprepitant, bosutinib, doravirine, eliglustat, naloxegol, osimertinib, tofacitinib). For these models, called the “Predictive Performance Subset,” the following were assessed: (i) the data available for PBPK model development, (ii) the workflow ⁵ followed, (iii) model limitations that could reduce its mechanistic basis, for example, fitted parameters, and (iv) whether any adjustment of the probe substrate or the rifampin model was needed before the rifampin DDI simulation was performed in initial model development.

Simulations

Simulations were conducted in the Simcyp Simulator software Version 20 (Certara, Radnor, PA) using the built‐in “Healthy Volunteer” population or “Cancer Patient” population, dependent on the relevant clinical DDI study or recapitulated modeling analysis. The cancer patient population incorporates differences in plasma albumin levels, alpha‐1‐acid glycoprotein levels, and glomerular filtration rate, but not in gut or liver CYP enzymes when compared with the healthy subject population. ³⁴ Simulation trial designs were set up to match the clinical DDI trial pertaining to doses, number of doses administered, times, and demographic information (percentage of females/males and age range) as closely as possible. The specific design of rifampin DDI simulations including the number of days of pretreatment and timing of initiation of the coadministered CYP3A substrate were selected to match the clinical DDI study being simulated (Table 2 ).

Table 2.

Details for rifampin DDI studies with CYP3A substrates included in this analysis ^a

Substrate (dose, regimen)	Study	Rifampin q.d. dose (pretreatment/total, days)	Stagger, hours	n	R AUC b	R Cmax b	Comments
IV dosing of substrate
Alfentanil	Multiplec	600 mg q.d. (NA)	~12 to 24	37	0.41	NA	Summary of studies found in Table S2
Midazolam	Multiplec	600 mg q.d. (≥ 7/≥ 7)	NA	97	0.46	NA	Summary of studies found in Table S2
Nifedipine (0.02 mg/kg SD)	Holtbecker et al. 35	600 mg q.d. (6–7/7)	~0 to ~24d	6	0.70b	NA	After 7 days rifampin, oral and iv nifedipine were applied on two consecutive days (random order)
PO dosing of substrate
Abemaciclib (200 mg SD)	Multidisciplinary review FDA	600 mg q.d. (5/14)	0	24	0.05 (0.04–0.06)	0.08 (0.07–0.09)
Alfentanil	Multiplec	600 mg q.d. (5 or 6/6)	~12 to 13	22	0.054	0.12	Summary of studies found in Table S2
Alprazolam (1 mg SD)	Schmider et al. 36	450 mg q.d. (3/4)	0d	4	0.12	0.64	R values calculated from reported mean PK parameters
Aprepitant (375 mg SD)	FDA review NDA 37	600 mg q.d. (8/14)	0d	12	0.09 (0.07–0.12)	0.38 (0.30–0.47)
Bosutinib (500 mg SD)	Abbas et al. 38	600 mg q.d. (6/10)	1	24	0.08 (0.07–0.09)	0.14 (0.12–0.16)
Doravirine (100 mg SD)	Yee et al. 39	600 mg q.d. (13/15)	0	10	0.12 (0.10–0.15)	0.43 (0.35–0.52)	Effect of single rifampin dose on doravirine PK was also studied; minimal effect
Eliglustat excluding CYP2D6 poor metabolizers (127 mg SD on day 1 and b.i.d. for 5 days)	Vu et al. 40	600 mg q.d. (5/6)	0d	19	0.15 (0.11–0.21)	0.16 (0.11–0.22)	A single dose of IV rifampin 600 mg on Day 1 followed by 600 mg PO rifampin q.d.
Eliglustat in CYP2D6 poor metabolizers (84 mg SD on day 1 and b.i.d. for 5 days)		600 mg q.d. (5/6)	0 h, 5d	6	0.04 (0.03–0.05)	0.05 (0.04–0.06)	A single dose of IV rifampin 600 mg on Day 1 followed by 600 mg PO rifampin q.d.
Ethinyl estradiol	Multiplec	600 mg q.d. (7–14/10–14)	~0 to ~12	48	0.354	0.64	Summary of studies found in Table S2
Everolimus (4 mg SD)	Kovarik et al. 41	600 mg q.d. (8/13)	0	12	0.37 (0.30–0.46)	0.42 (0.36–0.50)	Paper reported PK parameters based on blood concentrations, and thus the simulation was for blood concentrations
Ibrutinib (560 mg SD)	De Jong et al. 42	600 mg q.d. (7/10)	0d	18	0.152 (0.05–0.46)	0.079 (0.06–0.12)	Geometric mean (90%CI); AUCinf used for R AUC
Midazolam	Multiplec	600 mg q.d. (≥7/≥7)	−2 to ~12	185	0.101	0.139	Summary of studies found in Table S2
Naloxegol (25 mg SD)	Bui et al. 43	600 mg q.d. (9/10)	0	22	0.1 (0.1–0.1)	0.2 (0.2–0.3)	AUC and C max ratio are Geometric mean (95% CI)
Nifedipine (20 mg)	Holtbecker et al. 35	600 mg q.d. (6–7/7)	~0 to ~24	6	0.082b	NR	After 7 days rifampin, oral and iv nifedipine were applied on two consecutive days (random order)
Olaparib (300 mg SD)	Dirix et al. 44	600 mg q.d. (9/13)	0	22	0.13 (0.11–0.16)	0.29 (0.24–0.33)	Observed data in patients with advanced solid tumors
Osimertinib (80 mg q.d. for Days 1–49)	Vishwanathan 45	600 mg q.d. (−/21)	0	40	0.22 (0.20–0.24)	0.27 (0.24–0.30)	Observed data in NSCLC patients. Osimertinib was given for 77 days and rifampin was given for 21 days (days 29–49) with osimertinib at SS
Ribociclib (600 mg)	Samant et al. 46	600 mg q.d. (5/13)	0d	24	0.107 (0.0945–0.120)	0.190 (0.164–0.219)
Simvastatin (40 mg)	Multiplec	600 mg q.d. (5–8/5–9)	0–17	28	0.101	0.13	Summary of studies found in Table S2
Tofacitinib (30 mg)	Tse et al. 47	600 mg q.d. (7/8)	0	12	0.16 (0.14–0.18)	0.26 (0.23–0.31)
Triazolam (0.5 mg)	Villikka et al. 48	600 mg q.d. (5/5)	17	10	0.051	0.124
Venetoclax (200 mg SD)	Freise et al. 49	600 mg q.d. (9/13)	0	12	0.2	0.3	Geomean ratios based on venetoclax exposure with single dose of rifampin as reference

b.i.d., twice daily; IV, intravenous; NA, not applicable; NR, not reported; PO, oral; q.d., daily; SD, single dose.

^a For rifampin administration, the number of days or pretreatment (i.e., any rifampin doses administered before, even a short time such as an hour before, the CYP3A substrate is administered on the PK day to assess the DDI) as well as the total duration of administration is provided. Unless otherwise noted in the column with substrate information, studies are for a single dose of probe substrate. The stagger is the days between the last rifampin dose and CYP3A probe administration. A negative stagger means the rifampin continued after the CYP3A substrate was administered. RCmax values were not reported for IV administration.

^b Where R values were not reported, they were calculated as the ratio of summarized parameters. The range (when reported) is the 90% confidence interval unless otherwise noted in the comments. For IV PK, the RCmax was not typically reported and was not included here.

^c For some CYP3A substrates, data for multiple rifampin DDI studies were available. In this case, weighted values (see Eq. 3) for all studies dosed with 600 mg q.d. rifampin for ≥ 7 days are shown here and were used in the results figures. More details are found in the Methods and Supplementary Material Sections S3 and S4 .

^d Information on the dose stagger was ambiguous or not reported.

Model assessment

The analysis was performed by multiple working group members across different companies within the PBPK‐IWG. To ensure consistency a modeling workflow was developed and followed (Figure 1 ). First, it was ensured that each model could accurately simulate observed clinical PK. In some cases, the substrate model performance was assessed by simulating a clinical DDI study with a strong CYP3A4 inhibitor and evaluating against observed data (Table S1 ). After ensuring that a model was performing as expected, the rifampin DDI study was simulated.

Figure 1.

Workflow for model validation applied to all substrate models selected for analysis.

Single‐agent PK simulations consisted of 100 total subjects, 10 trials with 10 subjects each. For these types of simulations, summary statistics were derived from the pooled results across subjects. For each PK parameter, mean and standard deviation (SD) or geometric mean and geometric coefficient of variation (%GCV) were used to assess model performance, by comparison of summarized simulated PK parameters with observed data. For PK parameters, AUC_0–inf or AUCtau were used as appropriate whenever possible; otherwise, the AUC matched the reported data. For oral data, C _max was used, while for IV data, CL and Vss were compared with the observed data.

In cases where strong inhibitor studies were simulated as part of assessing substrate model performance, precipitant library models were used (i.e., ketoconazole and/or itraconazole). The dose and regimen were set to match the observed DDI study. DDI simulations consisted of 30–50 trials with 10 subjects each. Simulated trial statistics were compared with observed data as described for the rifampin DDI simulations below.

The goal of the model assessment was to confirm the models performed well in version 20 of Simcyp. In most cases, the observed data were adequately captured and there was no need for adjustment, but in some cases, changes were needed. For example, a model parameter could be missing in a publication or a model generated in an old software version might no longer be consistent with clinical data. In such instances, the models were adjusted (Table S1 ). Library models are regularly tested and updated so that further updates are unnecessary.

Rifampin DDI simulations

The Simcyp v20 rifampin model used in DDI simulations was the multiple‐dose (MD) rifampin model for repeat dosing with no food restriction. This model includes induction of CYP1A2, CYP2B6, CYP2C8, CYP2C9, CYP3A4, CYP3A5, and UGT1A1 and competitive inhibition of CYP2C8 and CYP3A4. Induction or inhibition of transporters including P‐gp was not incorporated. Details of the MD rifampin model are displayed in Table 3 .

Table 3.

Summary of salient features included in different PBPK models of rifampin available in different platforms

	PK‐Sim®	Simcyp™	GastroPlus®
Version	11.2a	20b	9.8
CYP3A4 induction
Indmax	9	16	15
IndC50 (μM)	0.34 (corrected for fraction unbound drug)	0.32	0.064 (corrected for fraction unbound drug)
Source	Meta‐analysis of published mRNA and activity data reported in vitro for primary human hepatocytes 18 , 50	“In vivo” induction EC50 and E max for rifampin derived from a change in 6β‐hydroxycortisol/cortisol after 600 mg q.d. rifampin for 14 days	Induction parameters were optimized using PBPK model to capture observed interaction with midazolam 51
Other pathways induced
Enzymes	CYP1A2, 2C8, 2E1, AADAC (auto)	CYP1A2, 2B6, 2C8, 2C9, 2C19; UGT1A1	CYP3A4 (auto), 2C8, 2C9; UGT1A1, 1A3(auto), 1A4, 2B7
Transporters	P‐gp (auto), OATP1B1 (auto)	–	P‐gp
Substrate pathways	AADAC, P‐gp, OATP1B1	Non‐specific clearance	CYP3A4, UGT1A3, OATP1B1, MRP2

^a PK‐Sim is part of the Open Systems Pharmacology (OSP) Suite Version 11.2.

^b Refers to the multiple‐dose (MD) rifampin model file (Rifampin‐MD) within Simcyp.

Rifampin DDI simulations consisted of 30–50 trials with 10 subjects each (Supplementary Materials Section S1 ). Simulated median, 5^th and 95^th percentiles (i.e., 90% prediction intervals (PIs)) of trial geometric mean ratios (GMRs, substrate + inducer/substrate alone) for AUC and C _max were compared with observed GMRs and 90% confidence intervals (CIs). The simulation GMRs for AUC and C _max for each trial were obtained directly from the output files. Median and 90% PIs of trial GMRs were calculated in Excel (Microsoft Corporation), with the 90% PI (prediction interval) determined using the equation: PERCENTILE (array, k), which returns the kth percentile of values in a range specified by the array, where k is 0.05 for the 5th percentile (low end of the PI) and 0.95 for the 95th percentile (high end of the PI).

Summarization of data for CYP3A substrates with multiple rifampin studies

For most of the CYP3A substrates, only a single rifampin DDI study was available based on searching in the University of Washington Drug Interaction Database (DIDB), now called the Certara DIDB®. But for alfentanil, midazolam, ethinyl estradiol, and simvastatin, multiple studies were available (Table S2 ). For midazolam and alfentanil, for both IV and PO administration, there were multiple DDI studies with rifampin. The weighted mean (WX) was calculated for the ratio of PK parameters with and without rifampin coadministered, that is, R values, from the available studies and used for comparison with simulated data as follows:

$$\mathrm{WX}=\frac{{\sum }_{j=1}^n{n}_j*x_j}{{\sum }_{j=1}^n{n}_j}$$ (3)

where n _j is the number of observations and x _j is the mean value from the jth study.

Data analysis

The agreement of simulated and observed DDIs was evaluated using forest plots constructed for the AUC and C _max ratios as well as the percent of induction captured, calculated as:

$$\frac{\left(1-\text{predicted ratio of change from baseline}\right)}{\left(1-\text{observed ratio of change from baseline}\right)}\times 100$$ (4)

In addition, the criteria proposed by Guest et al., which considers the bias introduced at low DDI magnitude, was applied. ⁵² The geometric mean‐fold error (GMFE) was calculated as:

$$\text{GMFE}={10}^{\left(\frac{\sum \left|\log \left(\frac{\text{predicted}R\text{value}}{\text{observed}R\text{value}}\right)\right|}{n}\right)}$$ (5)

with n = the number of predictions. The average fold error (AFE):

$$\mathrm{AFE}={10}^{\left(\frac{\sum \left|\log \left(\frac{\text{predicted}R\text{value}}{\text{observed}R\text{value}}\right)\right.}{n}\right)}$$ (6)

was used to assess whether predictions were generally underestimates or overestimates (i.e., to assess bias).

Static model equations were used as tools to illustrate the features of molecules in the Discussion. Eq. 7 shows the mechanistic static DDI model simplified to consider only induction:

$$\begin{array}{c}\frac{{\mathrm{AUC}}_{\text{with inducer}}}{{\mathrm{AUC}}_{\text{without inducer}}}=\left(\frac{1}{C_{\mathrm{g}}\times \left[1-F_{\mathrm{g}}\right]+F_{\mathrm{g}}}\right)\times \hfill \\ \left(\frac{1}{C_{\mathrm{h}}\times f_{\mathrm{m},\mathrm{CYP}3\mathrm{A}}+\left[1-f_{\mathrm{m},\mathrm{CYP}3\mathrm{A}}\right]}\right)\hfill \end{array}$$ (7)

where C is the extent of induction, which is dependent on site‐specific driving concentration and induction parameters for the inducer as well as the maximum induction effect (E _max) and the concentration causing half‐maximal effect (EC₅₀), f _m,CYP3A is the fraction of systemic clearance of the substrate mediated by CYP3A, h denotes liver, and g denotes gut. ⁸ The second term in Eq. 7 accounts for changes in both hepatic availability, Fh, and systemic CL_int. ⁵³

RESULTS

Simcyp CYP3A substrate models

The recommended workflow for a CYP3A substrate, previously published, was used to assess whether CYP3A substrate models could be included in this analysis. ⁵ The modeling strategy depicted in Figure 1 was used to test the performance of the substrate model prior to using it to simulate an induction DDI with rifampin. This strategy was driven by the common practice that the performance of the substrate model should be validated and able to describe the available clinical data. All CYP3A substrate models were confirmed to perform sufficiently to proceed with simulating the DDIs with rifampin (Table S1 ).

Twenty substrate models were identified with sufficient data to proceed with the induction DDI evaluation. Substrates with a range of f _m,CYP3A4 (0.086–1.0), Fg (0.11–1) and Fh (0.09–0.96) were included. Details for each CYP3A substrate (e.g., whether a compound was considered Tier 1 or Tier 2) and selected results (e.g., F, Fa, Fg, and Fh with and without rifampin coadministration) are tabulated in Table 1 .

Rifampin model

Various published PBPK models were available for rifampin using different platforms. A comparison of the salient features of the rifampin PBPK model in Simcyp™, GastroPlus®, and PK‐Sim® is given in Table 3 . In the Simcyp MD rifampin model for induction, a first‐order absorption model is applied with fraction absorbed (Fa) and rate of absorption (ka) predicted from in vitro permeability data, and fit to observed data. ¹⁰ , ⁵⁴ A minimal PBPK model is used for distribution where all organs aside from the liver and intestine are combined. The Vss was obtained from a meta‐analysis of reported clinical IV PK data. ³¹ , ⁵⁵ , ⁵⁶ , ⁵⁷ , ⁵⁸ , ⁵⁹ While not mechanistically incorporated into the model, the autoinduction of rifampin clearance is accounted for by using the observed clearance from PK data on day 7 following repeated once‐daily i.v. infusions of 600 mg rifampin. Therefore, the MD model will capture the PK at a steady state, but it cannot describe single‐dose rifampin PK. ⁵⁷ Renal clearance was obtained from a meta‐analysis of observed clinical data. ⁶⁰ , ⁶¹ , ⁶²

With the first‐order model, the driving concentration for the induction in the gut is the portal vein concentration (fu_G × C_pv, where fu_G is the fraction unbound in the enterocyte; with fu_G = 1). The driving concentration for induction in the liver is the unbound plasma concentration. Induction Ind_max (maximal fold induction, E _max + 1) and IndC₅₀ (the concentration in μM that yields half of the E _max) for CYP3A4 and CYP3A5 were derived as described in Almond et al. ¹⁷ Briefly, in vivo biomarker data, specifically the ratio of 6β‐hydroxycortisol to cortisol, following multiple oral dosing of rifampin (600 mg q.d. for 14 days) ⁶³ and rifampin PK data ⁶⁴ were used to obtain an Ind_max of 8‐fold and IndC₅₀ of 0.32 μM. The Ind_max value was then increased to 16‐fold to better recover observed DDIs with a set of CYP3A substrates after IV and oral administration (alfentanil, alprazolam, midazolam, nifedipine, simvastatin, and triazolam). Competitive inhibition (K _i ) values for CYP3A4 and CYP2C8 obtained from in vitro measurements are included in the model. ⁹

The enzyme turnover rate (k _deg) in the liver for CYP3A4/5 was 0.0193 hour⁻¹. This value was selected based on the ability to predict midazolam clinical DDI magnitude with a set of time‐dependent inhibitors with inactivation parameters (K _I and k _inact) determined using standardized methods. ⁶⁵ A k _deg value of 0.03 hour⁻¹ was applied in the intestine due to the higher turnover rate of enterocytes. Combined with the Ind_max and IndC₅₀ in Simcyp's MD rifampin model, these k _deg values have been shown to capture the time course of rifampin's inductive effect on midazolam, as well as the return to baseline. ⁶⁶

PBPK‐DDI simulations of CYP3A substrates with rifampin

The simulated and observed C _max and AUC ratios of CYP3A substrates for rifampin coadministration are presented in Figure 2 . Results for individual studies for substrates with more than 1 rifampin DDI study, alfentanil, ethinyl estradiol, midazolam, and simvastatin are found in Table S2 and Figures S1 – S5 . Overall, the analysis includes 10 reported DDI results for CYP3A substrates administered IV and 34 for CYP3A substrates administered PO. Evaluation of the simulated vs. observed C _max and AUC ratio values according to the Guest criteria ⁵² (Figure 2 c and 2 d , respectively) showed that Tier 1 substrates were generally within the pre‐defined Guest criteria, with mainly Tier 2 substrates outside the “successful prediction” region. The only Tier 1 substrates that did not fall within the region were the alprazolam PO study (AUC ratio only; C _max ratio was within the region) and the nifedipine IV rifampin DDI study. Although the reason for the nifedipine mismatch (AUC ratio, AUCR, observed = 0.70 and simulated = 0.50, considered a mismatch based on the Guest criteria, which is more strict for lower R values) is not clear, limited PK time course may have been a factor (i.e., 6 hours for nifedipine and 3 hours for nifedipine + rifampin).

Figure 2.

Forest plots of simulated (black) and observed (gray) C _max (a) and AUC (b) ratios for the 20 CYP3A substrates presented in Table 1 , with P‐gp substrates demarcated. The simulated vs. observed C _max (c) and AUC (d) ratios were evaluated against the Guest criteria (dashed lines) and the 0.8–1.25 region (dotted lines) for Tier 1 (blue) and Tier 2 (orange) with circles for PO administration and triangles for IV administration, P‐gp substrates marked with a + symbol and the diagonal solid line showing the identity. In panels (a) and (b), the blue box denotes Tier 1 CYP3A substrates, while the orange box denotes Tier 2 substrates. For rifampin DDI studies with CYP3A substrates administered IV (alfentanil, midazolam, and nifedipine), the R values for C _max were not reported.

The percent of observed induction captured by the models, calculated using Eq. 4, is displayed in Figure 3 . The magnitude of induction on AUC was captured within 80–125% of the observed data for all Tier 1 substrates, with the exception of IV administration of nifedipine, which had an underestimated AUCR, and PO administration of alprazolam, which was not far out of the region (Figure 3 b ). The AUCR was underestimated for everolimus, a Tier 2 substrate, leading to a percent of induction captured higher than 125%. The impact of rifampin on the percent of induction captured for the C _max ratio, C _maxR, was within 80–125% for fewer compounds, with two Tier 1 and six Tier 2 substrates outside of the range (Figure 3 a ).

Figure 3.

Forest plots of the percent of induction recovered for C _max (Panel a) and AUC (Panel b) for the 20 CYP3A substrates listed in Table 1 . The blue box denotes the substrate models designated as Tier 1 and the orange box denotes substrate models designated as Tier 2. Substrates of P‐gp are demarcated. The percent of induction recovered was calculated using Eq. 4.

As expected, simulation AUCR and C _maxR values indicated strong induction (ratio < 0.2) for all orally administered Tier 1 substrates with f _m,CYP3A4 values > 0.85, while observed data for all but naloxegol C _maxR (0.24) were also below 0.2. In addition, simulations with the eliglustat model, which has a low f _m,CYP3A4, matched the observed strong induction by rifampin. Additional analysis was conducted to understand the factors involved in the strong induction of eliglustat despite the relatively low f _m,CYP3A4 (see Case Study).

An analysis was conducted to assess the agreement of the predicted and observed DDIs with rifampin PO DDI studies overall and for various subsets (e.g., Tier 1 or Tier 2). The subset categories for each molecule are included in Table 1 . The overall GMFE was 1.62 and 1.54 for AUCR and C _maxR, respectively, indicating good concordance of predicted and observed data (Table 4 ). Simulations were more accurate for Tier 1 than Tier 2, but Tier 2 also has reasonable accuracy (GMFE < 2). Simulations were more accurate for compounds that are not P‐gp substrates or that are P‐gp substrates but have high permeability and were less accurate for P‐gp substrates with low‐to‐moderate permeability, based on comparison of percent of predicted R values within the Guest limit and GMFE values (Table 4 ).

Table 4.

Statistical analysis of PO simulation performance based on substrate attributes ^a

Dataset	Number of substrates	% within 0.8–1.25 range AUCR/C maxR	% within Guest limit AUCR/C maxR	GMFE AUCR/C maxR	AFE AUCR/C maxR	% of induction captured within 0.8–1.25 range AUCR/C maxR
All substrates	20/19c	25.0/21.1	75.0/68.4	1.62/1.54	1.09/1.03	95/57.9
Tier 1 substrates	11/10c	36.4/30.0	90.9/100	1.47/1.35	1.12/0.98	90.9/80.0
Tier 2 substrates	9	11.1/11.1	55.6/33.3	1.82/1.78	1.06/1.09	88.9/33.3
Substrates used for the development of Qgut or rifampin model (first‐order absorption model)	6/5c	33.3/20.0	66.7/80.0	1.65/1.52	1.01/0.76	83.3/80.0
ADAM model substrates	5	0.0/20.0	40.0/40.0	1.99/1.76	0.96/1.15	80.0/40.0
Substrates not used for the development of Qgut or rifampin model (first‐order absorption model)	9	33.3/22.2	100/77.8	1.43/1.45	1.23/1.15	100/55.6
Predictive Performance Subsetb	8	25.0/25.0	100/85.7	1.48/1.46	1.27/1.21	100/62.5
Not substrates of P‐gp	6/5c	33.3/20.0	66.7/80.0	1.65/1.52	1.01/0.76	83.3/80.0
High permeability	6/5c	33.3/20.0	66.7/80.0	1.65/1.52	1.01/0.76	83.3/80.0
Low or moderate permeability	0	NA	NA	NA	NA	NA
Pgp substrates	14	21.4/21.4	78.6/64.3	1.61/1.55	1.13/1.15	92.9/50.0
High permeability	7	28.6/42.9	100/71.4	1.40/1.38	1.23/1.21	100/42.9
Low or moderate permeability	7	14.3/0.0	57.1/57.1	1.85/1.74	1.03/1.21	85.7/57.1

^a Calculations here were done for PO simulations using Eqs. 5 (GMFE) and 6 (AFE). For IV simulations, n = 3, GMFE = 1.28 and AFE = 1.02. No additional calculations were done due to the small number. The eliglustat CYP2D6 poor metabolizer data were excluded. For tofacitinib, the simulation with CYP2C19 induction included (Ind_max = 16) was used. These results are from R values shown in Figure 2 , except for the last column, which is based on % of induction captured (see Eq. 4, Figure 3 ).

^b The “predictive performance subset,” including abemaciclib, aprepitant, bosutinib, doravirine, eliglustat, naloxegol, osimertinib, and tofacitinib, includes compounds not used for the development of the Qgut model, with a complete data package and no disqualifying model limitations, that used the recommended workflow, and used the rifampin model without adjustment (see Methods).

^c One substrate, nifedipine, did not have an observed C _max ratio available.

The substrates with simulated R values relative to observed values within 0.8–1.25 was assessed, but overall not many of the CYP3A substrates fell into this range (25.0% for AUCR and 21.1% for C _maxR overall). This range may be too high a bar given the variability that may be seen in DDI studies. For example, for the 10 PO midazolam DDI studies with 600 mg q.d. rifampin included in this analysis with similar designs and at least 7 days of pretreatment, the AUCR values ranged from 0.04 to 0.158, and the C _maxR values ranged from 0.05 to 0.298 (Table S2 , Figure S4 ).

For Tier 1 substrates, 90.9% and 100% were within the Guest criteria limit for AUC and C _max ratios, respectively. ⁵² AFE values indicate little bias, that is, similar over‐ and underprediction. The Qgut model resulted in accurate rifampin DDI predictions even for compounds not used for Qgut model development. For the 5 CYP3A substrates using the ADAM model for absorption, simulations were less accurate. However, four of these compounds (everolimus, olaparib, ribociclib, and venetoclax) were P‐gp substrates with low‐to‐moderate permeability, that is, compounds where P‐gp is likely to play an important role. ⁶⁷ The inaccuracy for these four compounds may be at least in part because rifampin induces P‐gp, but the effect of rifampin on P‐gp was not incorporated in the PBPK modeling. But this assessment is complicated because the models were developed using strong CYP3A4 inhibitor study to confirm the f _m,CYP3A4, and strong CYP3A4 inhibitors (e.g., ketoconazole and itraconazole) are often also P‐gp inhibitors, which was not incorporated in the modeling either. Of these 4 compounds, the 2 with the lowest absorption, everolimus and venetoclax, had underestimated R values and the 2 with the highest absorption, olaparib and ribociclib, had overestimated R values.

The predictive performance subset, including substrates that incorporated first‐order absorption and the Qgut model, that were not used for Qgut or rifampin model development, and that had no disqualifying PBPK model limitations, followed the workflow described by Hariparsad et al., ⁵ and used the rifampin model without adjustment, was included in this statistical analysis. Importantly, none of these CYP3A substrates had clinical data allowing for direct estimation of Fg (e.g., no grapefruit juice study), and relied on the workflow approach using the strong CYP3A4 inhibitor clinical DDI study to validate the model including the Fg estimate. The performance of the models for these substrates was close to that of the models for substrates used for Qgut and rifampin model development, which all had clinical data allowing the determination of Fg. For the Predictive Performance Subset, seven compounds out of eight were within the Guest limit range of acceptable predictive accuracy for C _maxR, and 8 out of 8 for AUCR.

Impact of rifampin CYP3A4 K_i on underestimation of changes in C _max

As described, the MD rifampin model includes the reversible inhibition of CYP3A4 (K _i  = 18.5 μM). The CYP3A4 K _i was determined in human liver microsomes with midazolam as a probe substrate. ⁹ The single‐dose (SD) rifampin model in Simcyp was validated with a single‐dose study with rifampin where the midazolam AUC ratio was reported to be 1.2. ⁶⁸ To evaluate whether underprediction of the repeat‐dose rifampin C _max DDI ratio for ribociclib, osimertinib, olaparib, tofacitinib, and doravirine could be due to overestimation of the Fg due to first‐pass metabolism, inhibition of metabolism test simulations without the CYP3A4 inhibition K _i on ibrutinib (high CL_int, low Fg) and triazolam (higher Fg, lower CL_int) were conducted. As expected, including the K _i made a difference with ibrutinib but not triazolam (Table S3 ). With ibrutinib, however, the rifampin DDI was further overestimated when the K _i was not included.

Role of intestinal metabolism

The prediction of Fg for CYP3A substrates has been studied from multiple perspectives. ⁵⁹ Estimation of Fg separately from an overall bioavailability term is necessary given that driving concentrations of inducer differ in the gut and the liver (Eq. 7). Drugs with Fg = 1 will not be inducible at the gut unless the drug is actually minorly extracted at the gut, but in this case, Fg cannot be easily distinguished from Fg = 1. Drugs included in the current analysis, such as alprazolam, doravirine, and osimertinib were predicted to have Fg of 0.99 or greater, which still enables induction to affect Fg. Drugs with Fg that is any value other than 1 are sensitive to be induced at the level of the gut.

In Simcyp models with first‐order absorption, the semi‐empirical “Qgut” model is used to estimate Fg. In development of the Qgut model, Yang et al. used a dataset of compounds where data directly informing Fg were available (e.g., grapefruit juice DDI studies, which with the right dose is thought to completely inhibit intestinal CYP3A4), including 6 of the 20 substrates included here (alfentanil, alprazolam, midazolam, nifedipine, simvastatin, triazolam). ²⁵ For drugs developed more recently, grapefruit juice studies are uncommon, and Fg is estimated indirectly.

Alternatively, a mechanistic absorption model such as the advanced dissolution and metabolism (ADAM) model, in which the drug is differentially exposed to CYP3A based on the distribution of enzymes in each segment, can be used to estimate Fg. Regardless of the absorption model, a DDI study with a strong CYP3A4 inhibitor helps refine parameters that define the sensitivity of a substrate to DDI at the gut vs. the liver. ⁵ Importantly, in the current analysis, rifampin DDI predictions were predicted well for compounds even when grapefruit juice studies are not available (e.g., see the Predictive Performance subset in Table 4 ).

Multiple parameters affect the sensitivity of a compound to induction in the gut. For all but one of the CYP3A probe substrates with a first‐order absorption model included in this dataset, the value for fu_G was 1 as recommended by Yang et al., ²⁵ based on the accuracy of the Fg predictions for this option compared with two others, estimating fu_G as the fraction unbound in plasma and the fraction unbound in blood. For these drugs, the effect of induction at the gut is driven primarily by how much fu_G*CL_u,int,G exceeds Qgut in the induced state (Figure S6 ). For compounds with absorption represented using the ADAM model, fu_G values were often lower than 1 (e.g., for ibrutinib, venetoclax, and olaparib).

When comparing the effect of rifampin across drugs, those with lower Fg in the baseline state showed larger relative decreases in Fg in the induced state (Table 1 ; Figure S6 ). Alprazolam and doravirine, with baseline Fg close to 1, were minimally induced at the gut, while drugs with lower baseline Fg were predicted to have Fg < 0.2 following induction by rifampin. Conversely, for the Tier 2 substrate osimertinib, Fg is predicted to be almost 1 (0.996) in the baseline state with the reduction to 0.10 in the induced state. This drug was modeled to be highly bound in enterocytes (fu_G = 0.00132 leading to a high Fg in the uninduced state) with low Qgut due to low permeability. ⁶⁹ This low fu_G value was required to reproduce the interaction with the strong CYP3A4 inhibitor itraconazole in an advanced cancer patient population. But this combination of factors makes the predicted induced fu_G*CL_u,int,G greatly exceed Qgut, leading osimertinib to be predicted to be the most sensitive of these 20 substrates to induction at the gut. For context, the predicted osimertinib AUCR and C _maxR values were higher than observed but met the Guest criteria for a successful prediction (simulated AUCR = 0.349 and C _maxR = 0.447, observed AUCR = 0.22 and C _maxR = 0.27). Sensitivity of Fg to low baseline CL_u,int,G is explored in a parameter sensitivity analysis, PSA, in Figure S6 .

Case studies

The following case studies are some of the more interesting examples to identify challenges and highlight important considerations for good predictions.

Case study 1: Strong DDI with rifampin despite low f _m,CYP3A4 for eliglustat

Eliglustat systemic clearance is highly dependent on CYP2D6 (f _m,CYP2D6: 0.81) with contributions from CYP3A4 (f _m,CYP3A4: 0.086) and renal elimination (fraction excreted 0.07) in CYP2D6 extensive metabolizers (EMs) following a single dose. The ability of the model to reproduce the increased effect of rifampin on eliglustat in CYP2D6 poor metabolizers where CYP2D6 does not extract the drug at the liver adds confidence that the model is well parameterized. CYP2D6 is not inducible and while there is a low CYP2D6 abundance in the gut, there is no evidence that CYP2D6 contributes to the metabolism of drugs in the intestine. Brexpiprazole and nebivolol models are two examples where gut input parameters are optimized to reduce dependence on CYP2D6 in the gut. ⁷⁰ , ⁷¹ For the current eliglustat model, the CYP2D6 abundance in the gut was set to 0. Eliglustat is a P‐gp substrate with high permeability, and so it is assumed in the modeling that P‐gp mediated DDI will not affect eliglustat disposition. As currently modeled, eliglustat f _m,CY3A4 is 0.086, Fh = 0.11, and Fg is 0.55 following a single dose in CYP2D6 extensive metabolizers (EMs), and the predicted rifampin AUCR is 0.14 vs. the observed AUCR of 0.15. Despite the limited dependence of systemic CL on CYP3A4 in CYP2D6 EMs, coadministration of eliglustat with rifampin reduced the exposure of eliglustat by 85%. Two features of eliglustat are responsible for this apparent disconnect. First, due to time‐dependent autoinhibition of CYP2D6, the dependence of systemic clearance on CYP3A4 increases with chronic dosing. Simulations with the current model suggest that at steady state, f _m,CYP3A4 for eliglustat increases to about 0.35 at steady state. Additionally, eliglustat also has low Fg due to CYP3A4‐mediated metabolism in the gut, therefore, this pre‐systemic metabolism is an elimination pathway. Regulatory guidance, including the ICH M12 guidance on DDI, ⁷² indicates the general need for sponsor companies to identify and characterize enzymes that contribute at least 25% to drug elimination. If elimination is pre‐systemic, it must still be characterized. Gut metabolism is potentially relevant for substrates of CYP3A4 and some other enzymes including UGTs and sulfotransferases. Even with a low contribution of systemic CYP3A4, a significant role in intestinal metabolism cannot be ruled out.

Another example of a drug with a low f _m,CYP3A4 and a significant DDI with rifampin is metoprolol, which is not included in the CYP3A substrates in the current study. Metoprolol is an old drug, and the contribution of CYP3A4 was possibly not initially appreciated, but it can illustrate how strong induction DDI can affect a high clearance drug with low f _m,CYP3A4, low Fh, and high Fg. Using the Simcyp v20 metoprolol model in a CYP2D6 EM population with no CYP abundance in the gut to simulate a single 100 mg dose of metoprolol in a 10 × 10 trial, the f _m,CY3A4 is 0.08, Fh = 0.51 and Fg is 0.99. The CYP3A4‐mediated metabolism inputs were informed by in vitro data. Following the dosing of rifampin (multiple dose) for 10 days with metoprolol dosed simultaneously on Day 11, f _m,CY3A4 is 0.32, Fh = 0.33, and Fg is 0.96 with a predicted AUC ratio of 0.73 vs. an observed AUC ratio of 0.67. ⁷³ Both eliglustat and metoprolol illustrate that a low f _m,CYP3A4 does not preclude induction by a strong CYP3A4 inducer and that current, comprehensive, mechanistic approaches can accurately reproduce induction DDIs even for drugs with < 25% contribution of CYP3A4.

Case study 2: Underprediction of rifampin and tofacitinib DDI

Both CYP3A4 (f _m,CYP3A4 = 0.52) and CYP2C19 (f _m,CYP2C19 = 0.17) metabolism are involved in the clearance of tofacitinib, as well as renal elimination (fe = 0.31). Although CYP2C19 contributes to the overall metabolic clearance of tofacitinib and is strongly induced by rifampin, the base multiple dose tofacitinib model in Simcyp v20 does not capture this pathway. An underprediction of the effect of rifampin was noted in the PBPK model with no overlapping CIs between predicted and observed AUC and C _max values. Sensitivity analysis to determine the impact of IndC₅₀ and Ind_max values on Fg, Fh, and AUCR indicated Fg is not sensitive to these parameters but Fh and AUCR are highly correlated. Since tofacitinib exhibits high passive permeability and high absorption, the role of P‐gp for the underprediction was ruled out. A recent publication sought to optimize the induction parameters for CYP2C19 using clinical interaction data and simulations in Simcyp v15. ⁴⁷ Incorporating the CYP2C19 induction in the model assuming the maximal induction equivalent to CYP3A4 (CYP2C19 Ind_max = 16) did not significantly improve the accuracy of the prediction. Therefore, increasing the CYP2C19 Ind_max to 20 was explored as a sensitivity analysis based on R values ≤ 0.1 observed for the effect of rifampin on CYP2C19‐sensitive substrate omeprazole indicating a relatively high Ind_max may be appropriate. ⁷⁴ , ⁷⁵ This higher Ind_max resulted in an increased prediction accuracy (Table S4 ). Based on the available evidence, it is possible that inaccuracy in the rifampin Ind_max value for CYP2C19 and/or the CYP2C19 enzyme turnover value contributed to the underprediction of magnitude of DDI. These results highlight the need for further understanding of induction mechanisms beyond CYP3A.

Case study 3: Potential role of P‐gp in underprediction of rifampin and ribociclib DDI

Ribociclib is mainly metabolized by CYP3A4 (f _m,CYP3A4 = ~0.7) and flavin‐containing monooxygenase 3 (f _m = ~0.16–0.30). ⁴⁶ An increase of AUC by 3.21‐fold and C _max by 1.67‐fold upon coadministration of a strong CYP3A4 inhibitor ritonavir can be rationalized with the in vitro f _m,CYP3A4 of ribociclib. PBPK modeling recovered the ritonavir DDI effects (AUCR: 2.84‐fold; C _maxR: 1.36‐fold) accurately. However, the magnitude of the rifampin DDI was underestimated by ~2.5‐fold (AUCR: 0.107 (observed) and 0.270 (predicted)). In addition to CYP3A4, ribociclib is also a substrate of P‐gp. The underprediction of rifampin induction could be due to intestinal P‐gp not being mechanistically included in the model. For orally administered drugs that are P‐gp and CYP3A4 substrates and where an intestinal first‐pass effect due to metabolism potentially occurs, the induction effect on intestinal P‐gp and CYP3A4 may both be needed in the model to accurately capture the rifampin DDI.

DISCUSSION

The IQ PBPK‐IWG began in 2020 with the aim to improve confidence in PBPK model predictions of induction. The group builds upon previous work conducted by the IQ Induction Working Group (IWG) and PBPK‐IWG which focused on in vitro aspects of CYP3A4 induction given the large amount of available clinical and in vitro data. ²⁰ , ⁷⁶ , ⁷⁷ , ⁷⁸ The first action for the PBPK‐IWG was to survey the industry for gaps and challenges and establish best practices. This included the proposal that PBPK models for CYP3A substrates which demonstrate independent validation of Fg and f _m and adequate recovery of clinical inhibition DDI could be used to waive or defer clinical DDI studies with inducers. ⁵ At the same time, concern with underprediction of rifampin DDIs has been raised by regulatory agencies, particularly when substrate models do not account for all inducible pathways. ⁷⁹ , ⁸⁰ , ⁸¹ Considering this concern, this analysis focused on establishing the prediction accuracy of PBPK modeling for predicting the effect of strong CYP3A inducer rifampin on CYP3A substrates with a range of f _m,CYP3A4 values.

Several considerations for PBPK modeling of induction have been highlighted including the need for substrate model validation with consistent lines of evidence including the agreement between the in vitro, clinical, and human mass balance data and knowledge of the f _m and Fg. The eliglustat case study and the metoprolol example demonstrate that drugs with high first‐pass extraction and low f _m,CYP3A4 may have measurable effects upon induction by a strong inducer. Additional questions to consider are whether additional inducible pathways contribute to the elimination of the substrate (i.e., CYPs, UGTs, P‐gp) and whether there are confounding factors, such as non‐linear PK or complex DDIs. The results of the analysis conducted here demonstrate that the current MD rifampin model can recapitulate clinical observations, particularly when the substrate model has been well‐defined.

The work here strengthens our confidence to predict rifampin DDIs with CYP3A substrates using PBPK modeling. These results may increase confidence for deferring or waiving a strong clinical induction study and informing labeling language. Additionally, a shift in clinical study planning might be considered. Instead of conducting a study with the strong inducer rifampin, and then simulating a DDI with a moderate CYP3A inducer, the opposite approach may be considered. This option may be particularly important considering the inability to use rifampin in healthy volunteer studies due to nitrosamines. Proposed strong inducer alternatives carbamazepine and phenytoin have several challenges (e.g., both are narrow therapeutic index drugs and carbamazepine requires a dose titration due to adverse events). ¹³ A well‐validated, robust rifampin model may support clinical development strategies where the strong inducer scenario is simulated and DDI studies are conducted with moderate inducers. In some cases, evaluating a moderate inducer may be preferred, for example, due to the clinical relevance of the moderate inducer or the need to conduct a study in a patient population.

Model credibility

While it is understood that several aspects are considered when assessing the utility of PBPK models for informing clinical study designs, for waiving studies, or for informing the label and that there are no defined “yes or no” criteria. Rather, the decision for acceptable use is informed from the totality of the evidence including safety, efficacy, and quality of the data package. Given the increasing scientific basis for PBPK‐DDI modeling, its impact may continue to increase. From a regulatory perspective, the assessment of model credibility for predicting the impact of a strong CYP3A inducer on a CYP3A substrate with a validated PBPK model may be increased from the current analysis. Therefore, it may also be possible that higher impact applications, such as using PBPK to guide dose adjustment even when no strong CYP3A induction study has been conducted, may be considered. Future work should assess the accuracy of PBPK predictions for additional strong and moderate CYP3A inducers, for example, efavirenz, and other inducible enzymes, for example, CYP2C19.

The assessment of the credibility of the Simcyp rifampin model is given in Table 5 . The context of use is the prediction of the effects of rifampin, a strong CYP3A inducer, on the C _max and AUC of CYP3A substrates (with well‐characterized disposition pathways without impactful contributions from other inducible pathways, that are not P‐gp substrates, or that are P‐gp substrates but have high permeability) to inform dose adjustments and coadministration restrictions. The CYP3A substrates for which the credibility assessment was defined based on model performance are assessed in Table 4 .

Table 5.

Credibility assessment

Validation and verificationa	Verification	Regardless of the platform chosen, users must ensure (in conjunction with the software vendor as applicable) confirmation of the correctness of the underlying equations and model structure for the intended purpose
	Models	For rifampin: published model of multiple‐dose rifampin for induction as utilized in Simcyp version 20 17
		For CYP3A substrates: workflow (Figure 1 ) applied to previously published models or models reproduced from publicly available NDA submissions and updated for the current effort
	Comparators	Simulations compared with observed rifampin DDIs for 42 clinical DDI studies that were reproduced using designs based on information available in specified references
	Assessment	Model performance was adequate, as given in Table 4 and Figures 2 and 3
	Applicability	Rifampin model captures observed DDIs for CYP3A substrates that have well‐characterized disposition pathways without impactful contributions from other inducible pathways, that are not P‐gp substrates, or that are P‐gp substrates but have high permeability
Model risk	Model Influence	Influence is high when the rifampin DDI simulation influences labeling in the absence of a clinical induction study with a strong inducer
		Influence is lower with the availability of a clinical study with another strong inducer (e.g., carbamazepine or phenytoin), or with a moderate inducerb
	Decision Consequence (object drug‐dependent)	Potential for serious adverse events or outcomes (for example, treatment failure) associated with object drug	Object remains effective and adverse events do not increase with induction	Object efficacy decreases and/or adverse events increase with induction
		Strong inducers will be allowed with or without a dose adjustment	Low risk	High risk
		Strong inducers will be contraindicated or avoiding use will be recommended	Low risk	Low risk
	Credibility	Demonstrated for the context of use ✓

^a Validation is determining the appropriate model structure and parameterization and demonstrating that the model is in good agreement with observed clinical data. ¹ Verification is ensuring that the equations are correctly implemented and solved by the software package.

^b For this assessment, conducting studies in cancer patients with strong inducers such as enzalutamide and apalutamide is not considered viable. The only inducer that can be administered to healthy volunteers in doses that produce strong induction is rifampin 600 mg once daily. For strong induction, carbamazepine doses of > 300 mg twice daily and phenytoin doses of > 300 mg/day are required, ¹³ , ¹⁴ but these doses are not recommended for administration to healthy volunteers. Lumacaftor 200 mg q.d. = not strong. Rifabutin 300 mg q.d. dose = not strong.

The current analysis was conducted in Simcyp due to the large number of available CYP3A substrate models. However, PBPK‐DDI simulations can be performed in other software packages. The utility of PK‐Sim in the Open Systems Pharmacology (OSP) Suite for simulating DDIs between rifampin, itraconazole, clarithromycin, midazolam, alfentanil, and digoxin has been demonstrated. ¹⁸ Also, GastroPlus has been shown to simulate the effect of rifampin on the PK of midazolam, nifedipine, and triazolam including assessment of the influence of time between substrate and inducer administration ¹⁹ as well as for rifampin and saxagliptin in patients with renal impairment. ⁸² Additional work will help to gain confidence in DDI predictions (build credibility) for the effect of rifampin on CYP3A substrates in these other platforms. Much of the recommendations in this assessment apply to DDI predictions performed using PBPK modeling regardless of platform.

The work here did not include CYP3A substrate models with high complexity in the mechanisms involved. Currently, such complexity (e.g., significant role of extrahepatic/nonCYP metabolism or transporter/enzyme interplay) may result in less confidence in the PBPK model. ⁵ Future work may include assessment of model predictions for compounds with more mechanistic complexity.

Role of intestinal P‐gp

Absorption for 15 substrates in the current analysis was described as a first‐order process. The first‐order absorption model is limited in terms of mechanistic considerations that can be incorporated, including intestinal transporters such as P‐gp. P‐gp substrates were included in the current analysis, and more of the CYP3A substrates were P‐gp substrates than not (Table 1 ). Although the impact of rifampin on P‐gp was not explicitly included in these models, the DDI prediction was good for many P‐gp substrates. Reasons for this observation may include high substrate permeability and solubility, and high intestinal drug concentration gradients exceeding P‐gp K _m values, saturating P‐gp efflux. ⁸³ Consequently, the relevance of P‐gp transport and its incorporation into PBPK modeling may not be important in every case.

In recent work, PBPK models were developed for CYP3A and P‐gp substrates including some substrates for both, and the effects of P‐gp were specifically incorporated in models using mechanistic in vitro data. ²² This work included the assessment of DDIs between rifampin and these substrates. Including the effect of rifampin induction of P‐gp empirically, using a 3.5‐fold increase of intestinal P‐gp abundance, resulted in more accurate predictions. The current work incorporated P‐gp substrates, but the effect of P‐gp was not explicitly included in the models. The rifampin repeat‐dose model, assessed as part of the current work, did not include an effect on P‐gp either. Therefore, rifampin coadministration did not have an effect on Fa for any simulations included here, even for P‐gp substrates.

The bosutinib model included in this analysis did not incorporate P‐gp explicitly. ⁸⁴ Models that explicitly incorporate P‐gp were considered out of scope. A later refined PBPK model for bosutinib incorporated newly available oral bioavailability data and explicitly included P‐gp with an empirical effect of ketoconazole and rifampin on P‐gp. ⁸⁴ This work indicated that the effect of rifampin on bosutinib PK involved not only induction of CYP3A4‐mediated metabolism but also induction of P‐gp and therefore absorption. Although the earlier bosutinib PBPK model included in this analysis did not explicitly incorporate P‐gp, the DDI simulations generated following the current workflow were still in reasonable agreement with observed rifampin DDI data. It is reassuring that following the proposed workflow resulted in reasonable simulation results despite limitations of the model.

Recommendations

Both the EMA and FDA have issued guidance outlining the requirements for reporting PBPK models ² , ³ and have published model evaluations during the regulatory review process. ¹ , ⁸⁵ , ⁸⁶ Similarly, industry members have described perspectives of PBPK platform qualification, model verification, and reporting formats with examples addressing various aspects of PBPK analyses intended for regulatory submissions and outlined the extent of qualification and/or verification required dependent on the impact of the simulations. ⁸⁷ These references provide valuable insight into model validation, in terms of the process of determining the degree to which a model or simulation represents the real world. For example, the EMA has emphasized the importance of PSA and uncertainty analysis, indicating that PSA should be conducted for key parameters (i.e., can change model outcome) as well as uncertain parameters. PSA can be critical in understanding the importance of key model parameters model (e.g., the impact of the fu_G value on simulations) and in gaining confidence in the model in scenarios where uncertainty may be high (e.g., when there is uncertainty in the f _m, multiple clearance mechanisms, or complex enzyme/transporter interplay). ⁵

For PBPK model building, it is recommended to begin with the “Bottom Up” approach and nonclinical ADME data to estimate Fg, f _m, and other parameters to start with a solid mechanistic basis for the model. Once clinical PK and other key data such as the human ADME study data and the strong CYP3A4 inhibitor study are available, the Fg and f _m should be confirmed, or refined if needed. PSA can be used to gain confidence in the model, for example, to understand how important intestinal metabolism may be to the DDI or to understand whether other inducible pathways might impact the DDI. Any mechanisms pertinent to the CYP3A substrate that rifampin might impact should be considered for inclusion in the model.

Intestinal metabolism by CYP3A is a key aspect of bioavailability that can contribute significantly to the DDI with rifampin for compounds that undergo intestinal metabolism, and care must be taken to parameterize this aspect of the model. Accurate estimation of baseline Fg is key to accurate predictions of the effects of inducers.

Predictions are most likely to be accurate for compounds that are not P‐gp substrates or that are P‐gp substrates but have high permeability. Along with PSA, alternative model structures should be considered, for example, ADAM model should be used instead of a first‐order absorption model for a P‐gp substrate with low permeability.

Regarding the rifampin model, several factors potentially impacting simulation results are whether to include the CYP3A4 inhibition K _i as well as other less characterized effects like induction of other enzymes and transporters of the substrate model. As previously noted, since CYP3A4 inhibition could have an impact on sensitive substrates with low Fg (see Table S3 ), it is advised to keep it as it reflects a non‐negligible inhibition in the intestine. This effect can however be minimized by staggering the co‐administered CYP3A substrate. For enzymes other than CYP3A4 or transporters, a sensitivity analysis around Simcyp induction parameters (especially Ind_max), or around literature values if not available in Simcyp, may help to reduce or understand the importance of the uncertainty. In addition, an appropriate choice of the substrate coadministration day is required to also ensure the steady‐state levels of all enzymes/transporters involved in induction.

In conclusion, PBPK modeling may be used to accurately predict the effect of rifampin on CYP3A substrates. Predictions are most likely to be accurate for compounds that are not P‐gp substrates or that are P‐gp substrates but have high permeability, while predictions tended to be less accurate for P‐gp substrates with low‐to‐moderate permeability. For such substrates, a model incorporating the impact of rifampin on P‐gp may be needed. The quality of a prediction relies on having a robust data package and understanding key drug disposition pathways and mechanisms involved for both the CYP3A substrate and the CYP3A inducer. The results presented here are for predicting the effect of rifampin on CYP3A substrates but the approach should generally apply to using PBPK modeling to predict induction by any drug.

FUNDING

No funding was received for this work.

CONFLICT OF INTEREST

The authors declared no competing interests for this work.

ETHICS STATEMENT

This work is a meta analysis of data from other publications. As such, no preclincial or clinical studies were conducted by the authors in animals or humans for this publication.

AUTHOR CONTRIBUTIONS

All authors wrote the manuscript; M.B.R., T.D.C., L.d.Z., D.R., M.E.D., K.S.T., Q.F., V.P.R., K.S., K.U., D.M., and J.R. designed the research; M.B.R., T.D.C., L.d.Z., M.E.D., K.S.T., Q.F., V.P.R., K.S., K.U., D.M., and J.R. performed the research; M.B.R., T.D.C., L.d.Z., D.R., D.M., and J.R. analyzed the data.

Supporting information

Supplementary Materials S1