PBPK Modeling for Regulatory Submissions

Physiologically based pharmacokinetic (PBPK) modeling is the mechanistic sibling of the pharmacometrics toolkit — where population pharmacokinetic (PopPK) modeling starts from observed clinical concentration data and fits a statistical description of exposure variability across a population, PBPK modeling starts from physiology itself and builds a concentration-time prediction from the ground up, before a single dose has necessarily been given to a human. That distinction is why PBPK sits in its own regulatory lane: FDA and EMA both treat a PBPK model as a piece of evidence that has to earn its credibility through an explicit, inspectable verification process, not as a black-box output a sponsor can simply assert. This guide covers how that credibility is actually built, what a submission has to contain, and the specific regulatory decisions a PBPK model is realistically able to influence today.

How a PBPK Model Is Actually Built

A PBPK model represents the body as a series of compartments corresponding to real organs and tissues — liver, kidney, gut, muscle, adipose tissue, and so on — connected by blood flow at physiologically realistic rates. Three categories of input feed the structure:

System parameters — organ volumes, blood flow rates, tissue composition, enzyme and transporter abundance by organ — drawn from physiology databases rather than from the drug being studied. These don’t change from one compound to the next.
Drug-specific parameters — physicochemical properties (solubility, permeability, pKa), plasma protein binding, and metabolic clearance data, typically generated in vitro from human liver microsomes, hepatocytes, or recombinant enzyme systems, then scaled up mathematically to a whole-body prediction.
Trial design parameters — dose, formulation, dosing schedule, and the demographic composition of the population being simulated.

Combining system and drug parameters this way is what the field calls a bottom-up approach: the prediction is built from independent physiological and in vitro building blocks rather than fitted backward from clinical data. A related middle-out approach blends the two — a bottom-up structure with select parameters adjusted so the model reproduces a limited set of observed clinical data — and is common in practice once early clinical PK data exists. Running the model against a virtual population — hundreds to thousands of simulated individuals whose physiological parameters are sampled from real demographic and physiological variability, rather than a single “average” patient — is what lets a PBPK model produce a predicted distribution of exposure, not just a single number.

Verification and Validation: Why a PBPK Model Has to Prove Itself First

Because a bottom-up prediction is built without fitting to the exact scenario it’s meant to predict, regulators require sponsors to demonstrate the model already gets *known* answers right before trusting it on an *unknown* question. This happens in two stages that are easy to conflate but are treated distinctly in both FDA and EMA guidance:

Verification — confirming the model’s mathematics and software implementation behave as intended (mass balance holds, the equations solve correctly), and that default system parameters reproduce established physiological benchmarks.
Validation — running the fully parameterized drug-specific model against real, independently observed clinical PK data the model was not tuned to (a different dose, population, or drug-interaction scenario than the one used to build it) and showing the prediction falls within a pre-specified acceptance range, commonly framed around a two-fold error criterion for exposure metrics such as AUC and Cmax.

A model that hasn’t cleared this bar for the specific question being asked isn’t considered “qualified” for that use, even if it performs well in some other context — qualification in PBPK modeling is tied to a context of use, not to the software platform or the drug in general. FDA’s guidance on the format and content of PBPK analyses (finalized in 2020, following an earlier draft) and EMA’s guideline on the qualification and reporting of PBPK modelling and simulation platforms (adopted by CHMP in 2018) both build their expectations around this verify-then-apply logic, though the exact acceptance criteria and documentation format differ in detail between the two agencies — sponsors running a global program should check each agency’s current guidance directly rather than assume one submission package satisfies both.

What a PBPK Submission Actually Has to Contain

FDA’s format-and-content expectations, in substance, ask a sponsor’s PBPK report to walk a reviewer through the same chain described above, documented rather than merely asserted:

The specific regulatory question the model is being used to answer (a defined context of use, not a general capability claim).
The model structure and the software platform and version used to build it.
Every input parameter, with its source cited — measured in-house, taken from the literature, or a software-default system parameter — so a reviewer can independently assess whether an input is appropriate.
Verification and validation results against independent clinical data, with the acceptance criteria stated in advance rather than chosen after seeing how well the model performed.
A sensitivity analysis identifying which input parameters the prediction is most sensitive to, since that’s where an uncertain input parameter poses the greatest risk to the conclusion.
The actual model files, not just a report describing them — reviewers at both agencies have said explicitly that they expect to be able to run the submitted model themselves, not take a sponsor’s summary result on faith.

What a PBPK Model Can Actually Change in a Development Program

The FAQ on CASRAI’s pharmacometrics guide already covers the headline use case — a verified PBPK model substituting for a dedicated clinical drug-drug interaction (DDI) study. That substitution is real but bounded: FDA has accepted PBPK-based DDI predictions in place of a clinical trial most consistently for enzyme-mediated interactions (CYP3A4 in particular) where the model has a track record of predicting similar interactions correctly, and continues to expect a clinical study for higher-risk interactions, narrow-therapeutic-index victim drugs, or mechanisms (certain transporter-mediated interactions in particular) where model performance is less established. Beyond DDI waivers, four other uses show up repeatedly in real submissions:

Dosing in organ impairment. Running a dedicated clinical PK study in patients with severe renal or hepatic impairment is often slow to enroll and ethically harder to justify for a small expected effect. A PBPK model that incorporates known physiological changes in impaired populations (reduced enzyme abundance, altered protein binding, changed blood flow) can support a starting dose recommendation, sometimes reducing the size or scope of the clinical study that’s still required.
Pediatric dose selection. Extrapolating an adult dose to children isn’t a simple weight-based scaling exercise, because organ maturation changes drug clearance nonlinearly with age. PBPK models built on pediatric physiology databases are one of the accepted tools under ICH E11A’s pediatric extrapolation framework for proposing a starting dose before a pediatric trial confirms it.
Absorption and formulation changes. Predicting how a change in formulation, or co-administration with an acid-reducing agent that alters gastric pH, will change absorption for a poorly soluble drug is a mechanistic question PBPK models are specifically built to answer, since gut physiology and pH-dependent solubility are represented explicitly in the model structure.
Supporting a labeling statement without a dedicated trial. Where a verified model already covers the mechanism in question, sponsors can sometimes support a labeling statement about an untested but mechanistically similar scenario directly from the model — the qualification bar above still applies, and this remains the exception rather than the default path.

PBPK vs. Population PK: Two Starting Points, Not Two Competing Methods

It’s worth being precise about a distinction the umbrella pharmacometrics guide only introduces briefly: PBPK and PopPK aren’t competing techniques answering the same question better or worse — they start from different information and answer different questions well. PopPK needs clinical concentration data to exist before it can say anything; its strength is describing how exposure actually varies across the specific population that generated the data, including sources of variability (genotype, organ function, drug interactions observed in the dataset) that a bottom-up model would have to be told about in advance to represent. PBPK needs no clinical data at all to produce a first prediction; its strength is answering “what would happen in a scenario we haven’t tested yet” — a different dose, population, or co-administered drug — which PopPK cannot do without collecting new data in exactly that scenario. Many development programs use both: a bottom-up PBPK model early, refined toward a middle-out model as early clinical data arrives, alongside a separate PopPK analysis once enough trial data exists to fit one directly.

Frequently Asked Questions

Does FDA require PBPK models to be built in a specific software platform?

No. FDA’s guidance is platform-agnostic and focuses on what the submission must document (structure, inputs, verification, sensitivity analysis, and the model files themselves) rather than mandating a specific tool. In practice, most sponsor submissions use one of a small number of established commercial or open-source platforms — Certara’s Simcyp Simulator, Simulations Plus’s GastroPlus, and the open-source PK-Sim/MoBi suite are the ones most commonly cited in the regulatory PBPK literature — because reviewers have accumulated familiarity with how each represents physiology, not because any one is formally required.

What’s the practical difference between “verification” and “validation” here?

Verification checks that the model and software are internally correct — the math solves right and default physiology matches known benchmarks. Validation checks that the fully drug-specific model correctly predicts real clinical outcomes it wasn’t built to fit. A model can be verified without being validated for a given use, and reviewers evaluate the two separately.

Can a PBPK model alone support a regulatory decision with zero clinical data?

Rarely, and only within a narrow, pre-defined context of use where the model has strong validation evidence for a closely analogous scenario. Agencies have been explicit that PBPK modeling is generally used to reduce, refine, or waive a specific study — not to replace clinical evidence across a development program wholesale.

Is PBPK modeling required for every IND or NDA?

No. It’s one tool within Model-Informed Drug Development that a sponsor chooses to apply where a mechanistic prediction can answer a specific regulatory question more efficiently than a dedicated trial — not a default requirement of every submission.