Why Test a Dose the Model Has Not Seen?
A PBPK model calibrated at one or more doses can reproduce those data yet still make unsupported predictions beyond them. Dose extrapolation asks a focused question: with the calibrated model and parameter values left unchanged, does the model predict concentrations at a dose excluded from calibration? This differs from asking whether a model can fit the new dose after adjustment. Refitting would use the test data to revise the model and would no longer test its original dose prediction.
The central comparison is between the predicted and observed concentration-time behavior at the excluded dose, considered in relation to dose. For a linear system with dose-proportional drug input, increasing the dose by a factor should increase concentrations, AUC, and Cmax by that same factor, while the dose-normalized profile should remain unchanged. Departures can reveal a failed assumption, but do not by themselves identify its cause.
The earlier material on calibration versus independent validation establishes why data reserved for testing must not be used to tune the model. Here the test is specifically across dose: the calibration dose or doses are kept separate from the dose selected to examine extrapolation.
What Dose Proportionality Requires
These comparisons require compatible measurements: the same analyte and concentration units, comparable sampling schedules and AUC definitions, and a consistent regimen and context. The dose must be defined on the same basis at each level (for example, the same active moiety). A change in formulation, route, or other conditions complicates interpretation because it can change drug input; those are distinct prediction questions.
Dose proportionality follows from linearity, not from the PBPK label. A model has linear dose response when its governing equations are linear in the relevant amounts and concentrations, its coefficients do not change with dose, and the input at dose \(D\) is \(D\) times the input at a reference dose. This generally requires fixed physiological conditions and dose-independent processes such as first-order elimination and unsaturated binding or transport. Saturable metabolism, transport, or binding can violate these conditions. So can a dose-dependent change in absorption or bioavailability, even if disposition after drug enters the systemic circulation is linear.
A short derivation clarifies the prediction. Let \(\mathbf{x}(t;D)\) be a vector of amounts in model compartments, let \(A\) be a fixed matrix of linear transfer and elimination terms, and let \(\mathbf{b}u(t;D)\) represent drug input. With zero initial amounts and input proportional to dose, \(u(t;D)=D u(t;1)\), the model satisfies:
The solution of this linear system is:
If concentration is a fixed linear function of compartment amounts, \(C(t;D)=\mathbf{q}^{\mathsf T}\mathbf{x}(t;D)\), then \(C(t;D)=D C(t;1)\). Integrating over the same time interval gives \(\mathrm{AUC}(D)=D\,\mathrm{AUC}(1)\). For \(D>0\), taking the maximum over time also gives \(\mathrm{Cmax}(D)=D\,\mathrm{Cmax}(1)\), because multiplying an entire profile by a positive constant multiplies its maximum by that constant. Each conclusion depends on the stated linearity and input assumptions; it is not guaranteed merely because a model was calibrated successfully.
A Practical Comparison at an Excluded Dose
Keep the calibrated model, parameter set, and prediction protocol fixed. Generate the prediction at the test dose before using its observations to revise anything. Then compare observed and predicted profiles, as well as AUC and Cmax, with results from the calibration dose. Dose-normalizing means dividing concentrations by dose, AUC by dose, and Cmax by dose. The units must be retained: if concentration is in mg/L and dose in mg, concentration per dose has units of L\(^{-1}\).
Record which dose or doses were used for calibration and which dose is excluded from calibration. Keep the relevant regimen and measurement definitions comparable.
Run the model at the excluded dose without changing its calibrated parameters in response to the test-dose observations.
Inspect concentrations over time and compare dose-normalized profiles. A mismatch in shape can be obscured by comparing only one summary value.
Calculate each endpoint using consistent definitions, then examine the endpoint per dose at calibration and test doses. Interpret discrepancies in light of the full profile and the model assumptions.
Worked Example: A Nearly Dose-Proportional Extrapolation
Suppose a hypothetical PBPK model was calibrated at 100 mg. Its predicted concentrations at selected times are 4.0 mg/L at 1 hour, 8.0 mg/L at 2 hours, and 5.0 mg/L at 6 hours. The predicted AUC over the specified interval is 120 mg·h/L and predicted Cmax is 8.0 mg/L. The 200 mg dose is excluded from calibration. With all parameters frozen, the model predicts concentrations of 8.0, 16.0, and 10.0 mg/L at those times, AUC of 240 mg·h/L, and Cmax of 16.0 mg/L.
Hypothetical observations at 200 mg are 8.4, 15.2, and 8.8 mg/L at 1, 2, and 6 hours, with AUC 228 mg·h/L and Cmax 15.2 mg/L. Compare the dose-normalized observations with the 100 mg calibration-dose predictions:
For AUC, the model predicts \(240/200=1.20\) mg·h/L per mg at 200 mg, exactly \(120/100=1.20\) mg·h/L per mg at 100 mg. The observed value at 200 mg is \(228/200=1.14\) mg·h/L per mg. For Cmax, the predicted normalized value is \(16.0/200=0.0800\) mg/L per mg, while the observed value is \(15.2/200=0.0760\) mg/L per mg; the 100 mg value is \(8.0/100=0.0800\) mg/L per mg.
The observed profile and endpoints are close to the dose-proportional prediction in this hypothetical example, with some differences, particularly at the later selected time. Whether the differences are acceptable depends on the prespecified purpose and evaluation criteria; these calculations alone do not establish that the model is adequate in every context.
Reading Patterns Without Overstating Their Meaning
The pattern across time and endpoints is often more informative than a single discrepancy. If observed dose-normalized concentrations are consistently higher at the test dose, and AUC per dose is also higher, the data are inconsistent with the model’s dose-proportional prediction. Capacity-limited elimination is one possible explanation: as dose rises, clearance may fall, causing exposure to rise more than proportionally. Dose-dependent bioavailability or another input change can produce a similar pattern. The comparison detects a problem with the predicted dose response; it does not distinguish these mechanisms on its own.
A discrepancy mainly in Cmax, with AUC per dose and much of the profile agreeing, has a different interpretation. Cmax depends strongly on the concentration-time shape around the peak and on sampling density. A shifted or sharper peak may alter Cmax without a comparable change in total exposure. This pattern calls for examination of the predicted and observed profiles, sampling times, and input assumptions; it is not, by itself, proof of nonlinear clearance.
AUC and Cmax are complementary summaries, not substitutes for the profile. AUC summarizes exposure over its specified interval but can conceal when concentrations differ. Cmax describes the peak but says little about the tail. Dose-normalized profiles expose time-varying differences that either summary can miss. Conversely, a few observed sampling times may not characterize a profile well, especially near the peak or at late times.
Worked Example: Supra-Proportional AUC at the Higher Dose
In a hypothetical study, the model was calibrated at 100 mg, where the observed AUC is 60 mg·h/L and Cmax is 6.0 mg/L. The frozen linear dose-scaling prediction for the excluded 300 mg dose is AUC \(=3(60)=180\) mg·h/L and Cmax \(=3(6.0)=18.0\) mg/L. Suppose the hypothetical test-dose observations instead give AUC 330 mg·h/L and Cmax 21.0 mg/L.
The observed AUC per dose is \(330/300=1.10\) mg·h/L per mg, compared with \(60/100=0.600\) mg·h/L per mg at the calibration dose. The observed AUC is \(330/180=1.83\) times the linear prediction. Cmax per dose is \(21.0/300=0.0700\) mg/L per mg, compared with \(6.0/100=0.0600\) mg/L per mg at calibration; observed Cmax is \(21.0/18.0=1.17\) times its linear prediction.
Both observed endpoints are greater than their dose-proportional predictions, especially AUC. This is a clear dose-response discrepancy for the hypothetical data. Saturable elimination is a plausible explanation, but the result cannot establish that mechanism: dose-dependent input or another mismatch in conditions could also contribute. The frozen prediction should be reported as the test, rather than replaced by a refitted prediction at 300 mg.
Worked Example: A Cmax Difference With Similar AUC
A hypothetical model calibrated at 50 mg predicts AUC 40 mg·h/L and Cmax 5.0 mg/L. At the excluded 100 mg dose, its frozen dose-proportional predictions are AUC 80 mg·h/L and Cmax 10.0 mg/L. Suppose observations at 100 mg yield AUC 82 mg·h/L and Cmax 13.0 mg/L.
Observed AUC per dose is \(82/100=0.820\) mg·h/L per mg, compared with \(40/50=0.800\) mg·h/L per mg at calibration. The observed AUC is \(82/80=1.025\) times its prediction. Observed Cmax per dose is \(13.0/100=0.130\) mg/L per mg, versus \(5.0/50=0.100\) mg/L per mg at calibration; observed Cmax is \(13.0/10.0=1.30\) times its prediction.
Here, total exposure is close to the prediction while the peak is higher. That combination suggests that the timing or shape of the concentration-time profile deserves attention before concluding that disposition is nonlinear. For example, altered input timing could change the peak while leaving AUC similar. This example illustrates why neither Cmax nor AUC alone is a complete dose-extrapolation assessment.
Assumptions, Diagnostics, and Alternatives
Before attributing a discrepancy to dose, check that the comparison genuinely isolates dose. The active substance, regimen, relevant physiological conditions, sampling and assay procedures, and endpoint definitions should be comparable. AUCs calculated over different intervals are not directly interchangeable. A sparse schedule can miss the true peak, so measured Cmax may be an observed maximum rather than a reliable estimate of the underlying peak. Solver convergence and extraction at observation times should also be adequate; numerical error can resemble a prediction failure.
A useful diagnostic is to examine whether discrepancies are broad and systematic across the profile or concentrated in a particular region. Consistent differences in dose-normalized concentrations and AUC suggest a scale-related mismatch. A difference limited to the peak suggests a shape or timing issue as well. If only a summary endpoint is available, the ability to distinguish these patterns is reduced. The test may also be uninformative if the excluded dose is not meaningfully separated from calibration doses or the observations are too imprecise to distinguish predictions.
The main alternative to a frozen PBPK prediction is an empirical dose-scaling description, such as estimating how AUC changes with dose from multiple observed doses. Such a description can summarize the observed trend, but it does not test the calibrated PBPK model’s prediction at an excluded dose if the test-dose data are used to estimate it. Refitting the PBPK model at the test dose can be useful for subsequent model development, but it answers a different question. Keep the original prediction and the subsequent revised model results distinct.
Dose extrapolation has limits. Agreement at one excluded dose supports the prediction only for the conditions and range actually examined; it does not prove dose proportionality at all doses or establish that every mechanism is correct. AUC and Cmax can agree while other parts of the profile differ. A mismatch indicates that the prediction or its assumptions need investigation, but does not uniquely diagnose the cause. Strong conclusions require an informative dose contrast, suitable observations, and a clear record of what was fixed before the test.
Check Your Understanding
Work through these questions, then compare your reasoning with the solutions.
- A model predicts AUC 45 mg·h/L at 75 mg. Under dose-proportional scaling, what AUC does it predict at 150 mg, and what is the predicted AUC per dose at each level?
- Why does a higher dose-normalized AUC at an excluded dose not prove that elimination is saturable?
- AUC per dose is similar at two doses, but observed Cmax per dose is higher at the test dose. Name one reason this pattern might occur without a comparable increase in total exposure.
- Why must the model parameters remain unchanged when evaluating the original dose-extrapolation prediction?
Solutions. (1) The dose doubles, so predicted AUC is \(45(150/75)=90\) mg·h/L. The normalized value is \(45/75=0.600\) and \(90/150=0.600\) mg·h/L per mg. (2) Dose-dependent input, differences in study conditions, or measurement and endpoint issues can also change AUC per dose; the observation identifies a departure from proportionality, not a unique mechanism. (3) A changed input rate or peak timing can change Cmax while leaving AUC similar. (4) Changing parameters using the test-dose observations uses those data to tune the prediction, so the result no longer tests the original frozen model.