Current Use of Physiologically Based Pharmacokinetic modeling in New Medicinal Product Approvals at EMA

Abstract

Physiologically Based Pharmacokinetic (PBPK) Models are routinely used in drug development and therefore appear frequently in marketing authorization applications (MAAs) to the European Medicines Agency (EMA). For a model to be a key source of evidence for a regulatory decision, it must be considered qualified for the intended use. Advice on the data expected to allow qualification of a PBPK model or platform is provided in the EMA Guideline on the reporting of PBPK modeling and simulation. The present study is an EMA review of the use of PBPK models in submitted MAAs in 2022 and 2023 focussing on the concept of qualification and the reasons why models were not considered qualified. A review of the 95 MAAs with a “full” legal basis approved during these years showed that 25 of them contained PBPK modeling. There were 65 proposed general areas of intended use for PBPK modeling identified across the applications, with the most common being a prediction of drug–drug interactions with enzymes or transporters (69%). Finally, this review showed that most of the models submitted in applications to EMA were not considered qualified for the intended use(s). The reasons identified for this are reported and the need for further EMA guidance, particularly around requirements for qualification of PBPK models, are discussed.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

PBPK modeling is an established tool in drug development and is regularly received in applications to EMA for approval of medicinal products submitted under a “full” (Article 8(3)) legal basis. This was seen in an EMA review of the use of PBPK modeling in regulatory submissions in 2016. Since then, in 2018 the EMA Guideline on the reporting of PBPK modeling and simulation was published.

• WHAT QUESTION DID THIS STUDY ADDRESS?

This study evaluated the use of PBPK models in “full” applications for medicines approved by EMA in 2022 and 2023 to inform on current trends.

• WHAT THIS STUDY ADDS TO OUR KNOWLEDGE?

Approximately one out of four approved “full” applications contained PBPK modeling in 2022 and 2023 and 28% of these were considered qualified as per the EMA Guideline on the reporting of PBPK modeling and simulation. Additionally, the review looks in more detail at the reasoning behind cases where qualification was not achieved.

• HOW THIS MIGHT CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

This article provides insight into the current use of PBPK modeling in applications to EMA, particularly in the context of the EMA Guideline on the reporting of PBPK modeling and simulation, and the need for additional guidance.

Modeling and simulation are used to support various stages of drug development, a process known as Model‐Informed Drug Development (MIDD). ¹ As such, regulatory authorities expect to receive model‐based analyses as part of dossier submissions. To allow such modeling to be used efficiently, regulatory authorities need to increase their understanding and confidence in the use of these approaches. Achieving this is a priority for the European Medicines Agency (EMA), citing optimization of capabilities in modeling, simulation, and extrapolation as an aim in the EMA “Regulatory Science to 2025” strategy. ² Over the past decade, there has been growing interest in MIDD from regulators worldwide, ³ including the drafting of International Council for Harmonization (ICH) M15: Model‐informed Drug Development General Principles Guideline to encourage international harmonization on the utilization of MIDD in drug development and regulatory decision making. ⁴

The 2018 EMA Guideline on the reporting of physiologically based pharmacokinetic (PBPK) modeling and simulation (the EMA PBPK Guideline) defines a PBPK model as “a mathematical model that simulates the concentration of a drug over time in tissue(s) and blood, by taking into account the rate of the drug's absorption into the body, distribution in tissues, metabolism and excretion (ADME) on the basis of interplay between physiological, physicochemical and biochemical determinants.” ⁵ While PBPK modeling was mentioned in several existing EMA guidelines at that time, ⁶ , ⁷ this was the first to specifically provide advice on what to include in a PBPK modeling report.

A retrospective study that evaluated the use of PBPK models in submissions to the EMA prior to the 2018 guideline showed an increasing trend in their use over the period 2004–2015 inclusive. ⁸ The submitted models were used to predict specific aspects of the pharmacokinetics of a drug candidate, from drug–drug interactions (DDIs) to absorption and bioavailability to pharmacokinetics in special populations. DDI studies were the most common use for PBPK models, particularly interactions either as a victim or perpetrator with enzymes such as cytochrome P450 enzymes (CYP450s) and Uridine 5′‐diphospho‐glucuronosyltransferases (UGTs).

PBPK models have the potential to complement or replace clinical pharmacokinetic studies (i.e., study waivers may be applied), to support decision making about the benefit/risk ratio and inform the Summary of Product Characteristics (SmPC) of a medicinal product. This can offer an attractive use of resources for applicants and provide the opportunity for a more holistic evidence base for regulatory assessment. If a model is the key source of information to be reflected in the product information (SmPC), the EMA PBPK Guideline considers the model to be of “high impact.” Models which contribute evidence to be assessed in combination with clinical data are considered “moderate impact.” Finally, models used for descriptive reasons (e.g., mechanistic understanding) in cases where a full clinical PK study is conducted are considered “low impact.” ⁵ , ⁹

For a model to be considered sufficiently robust, the terminology used by EMA since the publication of the 2018 guideline is that the model must be qualified for the intended use. The term “intended use” describes what information the model is seeking to predict, for example, in the case of a high‐impact model, which study it is intended to replace. To achieve this, a model needs to be built, then verified (confirmed that the model is correct from a platform/equations perspective), subsequently validated (the model's ability to describe the observed data demonstrated), and finally evaluated with respect to its ability to address the question of interest. Based on these steps, assessors will consider the model qualified or otherwise for its intended use(s). This qualification determines the extent to which a model is considered as evidence within the benefit/risk assessment. In practice, this means that for high‐impact models, qualification allows the results to inform prescribing information in the SmPC.

Qualification for an intended use may either be sought in a regulatory submission where the PBPK model is assessed or via a scientific advice qualification procedure to obtain a positive qualification opinion from the Committee for Medicinal Products for Human Use (CHMP) as a “novel methodology for drug development.” ¹⁰ The guideline also mentions the possibility of qualification supported by other sources of data such as peer‐reviewed literature. To date, qualification has only been obtained via assessment of models included as confidential data in regulatory submissions. Applicants may consider this route of qualification to require less time and resources than CHMP qualification, with only context‐specific data to be provided. In addition, there is a fee for CHMP qualification. However, this impacts transparency on the qualification of models (limited to the information within the public Assessment Reports), whereas if a PBPK model was qualified by a CHMP qualification procedure, the positive opinion would be published in detail on the EMA website.

There is no doubt that PBPK models have potential for further efficiency and streamlining of clinical pharmacokinetic studies, however regulatory confidence in these models is key to their acceptance and progression within the context of drug development. By providing an update on the uses and acceptance of PBPK models in new drug approvals over 2022 and 2023, this article reflects on where EMA is in terms of such confidence and future steps that may improve regulatory confidence.

METHODS

A search was carried out for reference to PBPK models in the clinical pharmacology section of European public assessment reports (EPARs) of all approved drugs in 2022 and 2023. ¹¹ , ¹² Reference to PBPK modeling was only found in EPARs of drugs submitted as “full” applications (Article 8 (3) legal basis (complete quality, non‐clinical, and clinical data included) dossiers). ¹³

A final list of procedures that included PBPK modeling was compiled (n = 25). The different Assessment Reports (ARs) (day 80, day 150, day 210) and list of questions/ outstanding issues (day 120, day 180) generated during the assessments were then searched in the EMA electronic records management system. For each procedure, an in‐depth manual review of the assessment was conducted, identifying the intended use of the PBPK model(s) and the outcome of the assessment of the model.

When identifying the intended use(s) of model(s), first the general areas of proposed intended uses were identified (e.g., DDI transporter) (Figure 2 , n = 65). These were further divided into specific proposed intended uses, at the level of the specific enzymes/transporters (e.g., DDI perpetrator P‐glycoprotein) (Figure 3 , n = 82).

Figure 2.

General areas of proposed intended uses of PBPK models in approved "full" MAAs in 2022 and 2023 (n = 65). Amount of proposed uses in each general area of intended use is indicated in brackets. Other: Bioequivalence (2), DDI acid reducing agent (1), Delayed gastric emptying (1), Effect of CYP3A4/3A5 polymorphisms in poor CYP2C19 metabolizers (1), Effect of food on DDI (1), Hepatotoxicity (1), Time until safe to be pregnant (1). CYP, cytochrome P450 enzyme; DDI, drug–drug interaction; perpetrator, where compound of interest acts as an inhibitor/inducer of an enzyme or transporter; victim, where compound is affected by an inhibitor/inducer of an enzyme/transporter.

Figure 3.

Use of PBPK models in MAAs in 2022 & 2023. Representation of number of (left to right) approved marketing authorization applications (MAAs) in 2022 and 2023 (166), approved ‘full’ MAAs in 2022 and 2023 (95), Approved ‘full’ MAAs in 2022 and 2023 containing PBPK modeling (25), proposed general intended uses (65) and proposed specific intended uses, that is, at the level of specific enzymes/transporters (82). CYP, cytochrome P450 enzyme; DDI, drug–drug interaction; PBPK, physiologically based pharmacokinetic modeling; perpetrator, where compound of interest acts as an inhibitor/inducer of an enzyme or transporter.

It was determined from the ARs whether each PBPK model had, or had not, been qualified and whether information from the model was used to inform the product information (SmPC). A model was classified by the authors as qualified if the assessment reports indicated acceptance of the model for the proposed intended use, not always directly using the word qualified. Phrases commonly seen were “endorsed,” “valid,” “appropriate,” “accepted.”

Additionally, reasons cited in ARs for PBPK models not being qualified were identified. The questions raised on Day 120 were particularly closely analyzed to gain insight into concerns raised during the initial assessment of a model. Note that the reasons in Table 1 are the author's interpretation of the ARs, which contained differing levels of detail depending on the procedure. These reasons were divided into 6 broader categories.

Table 1.

Proposed intended uses and outcomes of assessments of PBPK models in approved MAAs in 2022 and 2023

INN	Intended use(s)	Qualified or otherwise	Reason(s) for non‐qualification (where applicable)
Maralixibat chloride	DDI enzyme—perpetrator	Insufficient	Lack of relevant data to assess model's predictive performance
Maribavir	DDI enzyme—victim	Insufficient	Poor prediction of clinical data Lack of relevant data to assess model's predictive performance Insufficient justification of key model assumptions (incl. input parameters)
Mosunetuzumab	DDI enzyme—perpetrator	Insufficient	Insufficient justification of key model assumptions (incl. input parameters)
Voclosporin	DDI enzyme—victim	Insufficient	Poor prediction of clinical data Insufficient justification of key model assumptions (incl. input parameters)
	DDI transporter—perpetrator	Insufficient
	DDI transporter victim	Insufficient
Tirzepatidee	DDI delayed gastric emptying	Insufficient	Poor prediction of clinical data
Lutetium (177Lu) vipivotide tetraxetan	Renal impairment	Qualified
Mitapivat	DDI enzyme—victim	Insufficient	Concerns around PBPK model structure
	DDI—acid reducing agent	Insufficient	Poor prediction clinical data
Daridorexant	DDI enzyme—perpetrator	Insufficient	Concerns around PBPK model structure
	DDI transporter—perpetrator	Insufficient	Insufficient number of compounds in qualification dataset Insufficient justification of key model assumptions (incl. input parameters)
Lasmiditan	DDI enzyme—perpetrator	Insufficient	Insufficient number of compounds in qualification dataset
	DDI transporter—perpetrator	Insufficient	Insufficient number of compounds in qualification dataset
Asciminib	DDI enzyme—perpetrator	Insufficient	Lack of relevant data to assess model's predictive performance Concerns around PBPK model structure
	DDI enzyme—victim	Insufficient
	Hepatic impairment	Insufficient
	Renal impairment	Insufficient
	Food effect on DDIs	Qualified
	Absorption	Qualified
	DDI transporter—perpetrator	Qualified
Capmatinib	DDI enzyme—perpetrator	Qualified
	DDI enzyme—victim	Qualified
Atogepant	DDI transporter—perpetrator	Insufficient	Reason unclear from assessment report
	DDI transporter—victim	Insufficient
	DDI enzyme—victim	Insufficient	Poor prediction of clinical data
	Renal impairment	Qualified
Mavacamten	DDI enzyme—victim	Insufficient	Poor prediction of clinical data Insufficient justification of key model assumptions (incl. input parameters)
	DDI enzyme—perpetrator	Insufficient	CYP2B6 – Lack of relevant data to assess model's predictive performance CYP3A – Reason unclear from assessment report
	Renal impairment	Insufficient	Lack of relevant data to assess model's predictive performance
	Hepatic impairment	Insufficient	Insufficient justification of key model assumptions (incl. input parameters)
	Time until safe pregnancy concentration	Qualified
Glofitamab	DDI enzyme—perpetrator	Supportive	Lack of relevant data to assess model's predictive performance Insufficient justification of key model assumptions (incl. input parameters)
Adagrasib	DDI enzyme—perpetrator	Qualified only for CYP2D6 & CYP3A4.	CYP2B6 & CYP2C9 ‐ Lack of relevant data to assess model's predictive performance
	DDI enzyme—victim	Insufficient	Lack of relevant data to assess model's predictive performance Concerns around PBPK model structure
	DDI transporter—perpetrator	Insufficient	Lack of relevant data to assess model's predictive performance
	Hepatic impairment	Qualified
Ritlecitinib	PK Bioequivalence between dosage strengths (50 mg vs 100 mg)	Qualified
Futibatinib	DDI enzyme—victim	Insufficient	Concerns around PBPK model structure
	DDI transporter—perpetrator	Insufficient	Insufficient number of compounds in qualification dataset
Elacestrant	Hepatic impairment	Qualified for mild & moderate
	DDI enzyme—victim	Qualified
Omaveloxolone	DDI enzyme—victim	Insufficient	Poor prediction of clinical data Concerns around PBPK model structure
	DDI enzyme—perpetrator	Supportive	Reason unclear from assessment report
	DDI transporter—perpetrator	Supportive	Reason unclear from assessment report
	Absorption	Insufficient	Insufficient justification of key model assumptions (incl. input parameters)
Talquetamab	DDI enzyme—perpetrator	Qualified
Ivosidenib (Tibsovo/Tidhesco)	DDI enzyme—perpetrator	Qualified
	DDI enzyme—victim	Qualified
Quizartinib	DDI enzyme—perpetrator	Qualified
	DDI enzyme—victim	Insufficient	Concerns around PBPK model structure
Fezolinetant	DDI enzyme—victim	Qualified
	DDI transporter—perpetrator	Supportive	Insufficient number of compounds in qualification dataset
	Hepatotoxicity	Supportive	Lack of relevant data to assess model's predictive performance
Cabotegravir	DDI transporter—perpetrator	Supportive	Concerns around PBPK model structure
	DDI enzyme—victim	Insufficient	Concerns around PBPK model structure

Perpetrator, where compound of interest acts as an inhibitor/inducer of an enzyme or transporter. Victim, where compound is affected by an inhibitor/inducer of an enzyme/transporter.

CYP, cytochrome P450 enzyme; DDI, drug–drug interaction; INN, International Non‐proprietary Name; PBPK, physiologically based pharmacokinetic modeling; PK, pharmacokinetics.

Notably, the previous EMA study reported that PBPK modeling is not always represented in the public documents. ⁸ For this reason, for approved drugs in 2022, day 80 ARs and electronic common technical document (eCTD, summary of clinical pharmacology, module 2) for the 36 procedures with no mention of PBPK in the EPAR were searched for reference to PBPK models. The eCTD is the electronic submission of the dossier containing all data from the Applicant to be assessed by the regulatory agencies. This could contain data (e.g., modeling) that is not directly referenced in the ARs. Day 80 ARs are the initial ARs reflecting the independent assessment of the dossier by the rapporteur and co‐rapporteur teams and they are circulated to the National Competent Authorities of all Member States for their input in forming the pivotal Day 120 List of Questions. Subsequent questions to the applicant, responses from the Applicant, and further assessments mean day 80 ARs contain information that may not be mentioned in the eventual EPAR after approval of the drug. No additional models were mentioned in day 80 ARs that hadn't appeared in the EPARs. In one case, the term PBPK was present in day 80, 120, and 180 correspondences as a suggestion by the assessor however the Applicant never submitted one. One PBPK model was present in an eCTD which was not in the day 80 AR or EPAR, this was in relation to a QSP model which they used PBPK‐derived data to build.

RESULTS

Of the 48 marketing authorization applications (MAAs) approved under a “full” (Article 8 (3) legal basis) in 2022, 11 (23%) contained an explicit reference to PBPK modeling in the EPAR. ¹³  In 2023, this increased to 14 of 47 (30%). ¹² Three of the 25 approved “full” MAAs containing PBPK modeling were applications for monoclonal antibodies, and one was for a peptide drug.

The therapeutic areas of MAAs containing PBPK in 2022 and 2023 are shown in Figure 1 . The large proportion for oncology reflects the high overall number of oncology products approved. When normalized, PBPK was used in 36% of analyzed oncology MAAs.

Figure 1.

Therapeutic areas of approved "full" MAAs containing PBPK modeling in 2022 and 2023. Orange: approved in 2022. Blue: approved in 2023.

The general areas of proposed intended use(s) of the submitted model(s) are shown in Figure 2 . There were 65 proposed intended uses of PBPK models pertaining to the 14 general areas of intended use shown in Figure 2 . Some applications contained more than one PBPK model and most models submitted had more than one proposed intended use, ranging from two proposed uses for Lasmiditan to seven for Asciminib. As shown in Figure 2 , the prediction of DDIs with enzymes and transporters was the most common general area of intended use (69%).

Models were qualified for 18 of the 65 (28%) proposed intended uses. In cases where a model was intended to provide evidence for a statement in the SmPC, qualification resulted in either the model influencing a statement in the SmPC, or model results being directly represented in the SmPC (i.e., excluding uses such as bioequivalence uses). Of the three most common general areas of use, models predicting DDIs with enzymes (including all pathways, i.e., CYP and UGT enzymes) with the drug as a perpetrator had the highest rate of qualification (5/14, 36%), followed by DDIs with enzymes with the drug as a victim (3/16, 19%) and DDIs with transporters with the drug as a perpetrator (1/13, 8%).

In addition to cases where models were explicitly qualified in the ARs, some models were accepted as “supportive” for a statement in the SmPC (categorized as “supportive” in Table 1 ) despite the ARs indicating that the model was not qualified for the proposed intended use. This applied to six proposed intended uses of models across four applications. For example, for cabotegravir, modeling was considered supportive alongside in vitro data for a statement in the SmPC “caution is advised when co‐dosing with narrow therapeutic index OAT1/3 substrate medicinal products.” ¹⁴

The 65 proposed intended uses were further defined (e.g., at the level of a particular enzyme or transporter), into 82 specific proposed intended uses (as represented in Figure 3 ).

The most common specific intended uses are shown in Figure 4 . Predictions of DDIs with the CYP3A4 enzyme were the most common.

Figure 4.

Specific proposed intended uses of PBPK modeling in approved "full" MAAs in 2022 and 2023. Qualification rates per specific use are shown: Blue: Qualified, Orange: Supportive, Gray: Insufficient. Figure shows the most frequently proposed specific intended uses (seen in ≥ 3 cases). Total proposed specific intended uses (n = 82) and qualification rate is shown on the right‐hand side; perpetrator, where compound of interest acts as an inhibitor/inducer of an enzyme or transporter; victim, where compound is affected by an inhibitor/inducer of an enzyme/transporter.

Table S1 contains information from the EPARs on the assessment of PBPK models for the 11 approved “full” applications with reference to PBPK in 2022.

Table 1 summarizes the reason(s) for a model not being qualified based on the ARs. In the table, the term “insufficient” was applied to models that were not qualified in ARs and thus did not support a statement in the SmPC. Of note, Table 1 represents exclusively the intended uses that are specified in the EPAR in order not to disclose confidential information. The reasons for non‐qualification were identified based on the author's interpretation of the ARs and divided into 6 broad categories, ranked as follows:

Concerns around PBPK structure (14 cases)

Issues raised concerning the effect of the way the model was built on model function, such as the lack of incorporation of specific mechanisms including autoinhibition/autoinduction, the inclusion of intestinal enzyme/transporter activity and the incorporation of probe substrates. This can raise doubts regarding the model's ability to predict clinically relevant interactions.

Lack of relevant data to assess model's predictive performance (15 cases)

The qualification and verification data was not appropriate. This was often relatively poor clinical data due to non‐representative population characteristics (e.g., only healthy volunteer data) or issues with study design (e.g., non‐PK studies with sparse sampling, small sample size, or non‐relevant DDI being investigated). In other cases, relevant data (e.g., compound file) were not submitted.

Poor prediction of clinical data (11 cases)

The model was unable to accurately predict existing clinical data. For example, under‐ or overprediction of key PK parameters (e.g., T _max, C _max) or of the effect of inhibitors/inducers on PK of a drug. As stated in the EMA Guideline on the reporting of PBPK modeling and simulation, “The acceptance criteria (adequacy of prediction) for the closeness of the comparison of simulated and observed data depends on the regulatory impact.” ⁵

Insufficient justification of key assumptions (including input parameters) (12 cases)

Assumptions or input parameters, for example, relative contribution of certain enzymes in drug absorption and metabolism or enzyme affinity, were not justified (i.e., biologically plausible or valid). This was often based on uncertainty surrounding the source of the data (e.g., data from literature rather than from in‐house studies) or methods used to determine assumptions of the model (e.g., determining the relative contribution of enzymes).

Insufficient number of compounds in qualification dataset (7 cases)

The model's performance was not evaluated with enough compounds (e.g., substrates, inhibitors, inducers) during qualification to be considered robust. The EMA PBPK Guideline states that “for example, eight to ten compounds” is indicative of a sufficient number. However, this guidance was written considering the knowledge at that time for CYP3A4 and anticipating a rapidly evolving expansion of knowledge on DDIs involving other CYP enzymes. It is now acknowledged that 8–10 compounds might be challenging for several other CYP or UGT enzymes. Nevertheless, a sufficient number of compounds should be included to confirm accurate modeling prediction and to indicate the uncertainties in the model predictions.

Reasons unclear from assessment report (7 cases)

From the authors' perspective, a specific reason for the model not being qualified could not be decided upon ARs alone.

DISCUSSION

This study provides an update on the use of PBPK models in EMA‐approved “full” MAAs in 2022 and 2023. The results confirm the routine use of PBPK models in new drug development, reflected by 26% of approved MAAs containing PBPK in this period. Data from submitted PBPK models was accepted as evidence within MAAs (i.e., PBPK qualified or supportive) for 37% of proposed intended uses. The most common intended use for the submitted models was to predict DDIs with enzymes or transporters (69%), consistent with the previous review of PBPK modeling at EMA and with the review of PBPK modeling in applications to FDA. ⁸ , ¹⁵

The common use of PBPK modeling for DDI predictions is reflected by general guidance on the use of such models within the draft ICH M12 drug interaction guideline published in 2022. ¹⁶ This guidance should in turn give more confidence to Applicants submitting PBPK modeling for DDIs. 46% of proposed intended uses were the prediction of DDIs with enzymes. On a more granular level, when proposed specific intended uses were identified, 31% were for the prediction of DDIs with CYP3A4. The frequency of PBPK model submission for certain intended uses may allow identification of certain trends or general principles applied in EMA regulatory review practice and hence the publication of, for example, a Questions & Answers (Q & A) document similar to the 2023 EMA published Q & A on model‐based approaches for approval of alternative dosing regimens and routes of administration of anti‐PD‐1 and PD‐L1 monoclonal antibodies. ¹⁷ Such specific guidance for qualification requirements could be expected to increase both the use and acceptance of models for these intended uses.

Modeling to predict DDIs involving enzymes with the drug as a perpetrator had a relatively high level of qualification (36%) compared to the level across all intended uses (28%). Additionally, it was the proposed intended use for 2 of the 6 cases in this study where a model was accepted as “supportive” for a statement in the SmPC. The frequency of acceptance of PBPK models for this intended use may be an incentive for Applicants to apply for a CHMP qualification procedure for a platform for this use in the future. Although the Guideline on the reporting of PBPK models outlines several options for qualification as discussed in the introduction, the experience to date has been limited to model qualification within regulatory submissions. It is acknowledged that if a PBPK model receives a positive CHMP qualification opinion for an intended use via Scientific Advice, the publication of this opinion could make clearer the requirements for qualification and encourage future applicants.

Models were considered qualified for 28% of all proposed intended uses. For a further 9%, models were described as not qualified in the ARs but were accepted as supportive for a statement in the SmPC, alongside other data. This highlights that a model considered not qualified for the applicant's proposed intended use can still impact the SmPC to some extent. Whether a model can support a statement in the SmPC is dependent on multiple factors, with model impact and model risk being key.

As stated in the introduction, the impact of a model is related to whether there is other accompanying evidence submitted (e.g., in vitro / in vivo data) that suggests the same result. It is stated in the guidance that the “higher the impact, the greater the requirements on qualification of the PBPK platform.” However, independent of model impact, there are general principles EMA applies when assessing PBPK models. ¹⁸ On a case‐by‐case basis, assessors look at the intended use and assess drug models and the platform and related parameters (data sources, relevance, plausibility, etc.) for this intended use. Model credibility activities (model verification and validation activities), applicability, and uncertainty are also assessed. Predictive performance, taking into account the therapeutic window and the additional relevant data (if available), such as clinical study results, is also considered. It is acknowledged that a limitation of the current analysis is that the authors did not look in detail at these additional relevant data that may have been considered alongside the model to support statements in the SmPC. A further limitation is that the steps taken by assessors may not be systematically reflected in ARs or EPARs (see Table S1 ) so it is difficult to definitively report which steps were taken.

Issues raised by assessors when models were not qualified varied from case to case, defined by the authors into six broad categories. The most common was a lack of relevant data to assess the model's predictive performance, either in vivo data (due to target population or study design) or qualification data submitted by the Applicant (insufficient or not submitted). Next were concerns around the PBPK model structure. Finally, there were also a considerable number of concerns regarding the ability of models to predict known clinical data as well as concerns around key model assumptions such as input parameters and the source/justification of these assumptions. While these issues are addressed within the current Guideline, it is acknowledged that applicants could benefit from more streamlined and specific guidance, which EMA is currently working toward.

For seven proposed intended uses, the reason for the model not being considered qualified was not clear to the authors upon reading the ARs. In five of these instances, the intended use was to predict DDIs with transporters, one was to predict induction of an enzyme, and one to predict time‐dependent inhibition of an enzyme, all of which are regarded within the modeling and simulations network as less established uses of PBPK. This could be an implicit reason for non‐qualification, despite this being unclear in the ARs. In some cases, another potential explanation could be that the reason for non‐qualification can be deduced from reasons given for other proposed intended uses within the same application. For example, if there are concerns around the PBPK model structure cited as a reason for non‐qualification for one intended use, this can be interpreted as the reason for all intended uses of that model. However, if this is the case, it was not evident to the authors from analysis of the ARs.

Despite these possible explanations, it remains the case that there was variability in the level of detail given in ARs regarding model assessment and qualification. This is demonstrated in Table S1 which shows the EPAR text on PBPK models for the 2022 MAAs included in this analysis. A more standardized approach by assessors, using common terminologies for the assessment of PBPK models, could be helpful in reducing this. For example, there have been proposals to use a credibility matrix for models submitted within an application. ¹⁸ The implementation of such a matrix would standardize the level of detail that each AR contains where a PBPK model has been assessed. Additionally, it would make it easier for assessors to consider precedence with previous model assessments for similar intended uses. Further EMA guidance on the use of PBPK models in regulatory submissions could also help to improve the consistency and clarity of the information in ARs, including EPARS.

The outcome of a model not being considered qualified for its intended use during assessment can vary. In some cases, the Applicant was asked by the assessors to remove the information from the model from the SmPC (e.g., Mitapivat for a victim of moderate CYP3A4 inhibitor), in other cases quantitative components of statements proposed by the Applicant based on modeling were not included in the authorized SmPC resulting in a more general statement (e.g., for omaveloxolone the final wording in the SmPC was “Concomitant use of omaveloxolone with strong or moderate CYP3A4 inducers may significantly decrease the exposure of omaveloxolone, which may reduce the effectiveness of omaveloxolone”) ¹⁹ .

A limitation of the present study is that it reports on new drugs submitted under a particular legal basis that is, Article 8 (3) or a “full” application and looks only at the clinical pharmacology section of assessments. It is also the case that PBPK modeling is routinely also used, for example, in pediatric and non‐clinical applications submitted to EMA and this review likely under‐represents their use more generally.

While no reference was found in the EPARs of drugs approved as generics or biosimilars over the relevant timeframe, there were two instances where a model was used in the context of demonstrating bioequivalence within a new drug MAA, that is, one to support bridging between two formulations (capsule/tablet) and the other to support a strength biowaiver in the absence of a bioequivalence study. There has been interest in PBPK modeling to support biowaivers in the development of “complex generics” where there is recognition from a regulatory perspective of their potential ²⁰ , ²¹ . The use of PBPK modeling is also referred to in the draft ICH M13A on bioequivalence and as such a rise in PBPK use within generic applications in the following years is anticipated. ²²

This review provides insight into PBPK qualifications as per the EMA guideline. The results of this study will help shape future guidance on PBPK modeling. Given the apparent low proportion of models qualified, it could be that the current guideline is overly strict or not clear enough in defining the requirements for model qualification. It is acknowledged that in the current guidance, there is a lack of distinction between the data necessary for a model to be qualified within a given MAA and that for a model to receive a CHMP qualification opinion. This could be partially attributed to the use of the same terminology (“qualification”) for acceptance within an MAA and CHMP qualification. Future guidance should strive to further distinguish these processes and the requirements for each.

Furthermore, additional guidance is most likely to focus on particular uses of models, such as a Q&A for the prediction of DDIs as a perpetrator with CYP3A4. For any such guidance, there is a fine balance needed between prescriptive guidance to maintain clarity and flexible guidance to allow for continued applicability as technology progresses. This is a challenge associated with all MIDD which has previously been identified by regulators. ²³ To explore these issues further with stakeholders, EMA is planning a workshop on reporting and qualification of mechanistic models in 2025. This presents an opportunity to review the experience with the Guideline and to develop topics for further necessary guidance. Furthermore, a Q&A on reporting and qualification of PBPK models is mentioned as a high priority in the Methodology Working Party workplan 2022–2024 ²⁴ and EMA's participation in the drafting of the ICH M15 MIDD guideline will further contribute to clearer guidance for Applicants submitting PBPK guidelines to EMA. ²⁵

FUNDING

No funding was received for this work.

CONFLICTS OF INTEREST

The authors declared no competing interests for this work.

AUTHOR CONTRIBUTIONS

P.P., K.B., P.J.C., F.M.T., C.V., and E.M. wrote the manuscript. K.B., P.P., and F.M.T. designed the research. P.P. and P.J.C. performed the research. P.P., P.J.C., K.B., and F.M.T. analyzed the data.

DISCLAIMER

The views expressed in this article are the personal views of the authors and may not be understood or quoted as being made on behalf of or reflecting the position of the regulatory agencies or other organizations with which the authors are affiliated. The authors are employees of the European Medicines Agency or a National Competent Authority.

Supporting information

Table S1