Why Reduce a PBPK Parameter Set?
A PBPK model may contain many physiological and drug-specific inputs, but not every input has a material influence on every prediction of interest. A parameter reduction is a deliberate simplification of the model’s input set for a defined purpose: for example, fixing an input at a reference value so that calibration or repeated scenario analysis involves fewer free parameters. It is not a claim that the fixed physiological process is absent, nor that the parameter is unimportant for every drug, endpoint, or population.
The decision therefore depends on the outputs and scenarios the model must support. A parameter with little effect on plasma AUC may still affect peak concentration, a tissue concentration, or a different physiological population. The workflow here uses a hypothetical PBPK model for an orally administered drug. The aim is to assess whether renal clearance can be fixed at a reference value while retaining the ability to predict AUC, peak concentration, and an early liver concentration over a stated range of doses and physiology.
Define the Outputs and Analysis Domain
Suppose the hypothetical model has five inputs: intrinsic hepatic clearance \(CL_{\mathrm{int}}\), hepatic blood flow \(Q_h\), unbound fraction in plasma \(f_u\), absorption rate constant \(k_a\), and renal clearance \(CL_r\). The team selects three outputs before inspecting sensitivity results: plasma AUC from 0 to 24 hours (mg·h/L), plasma \(C_{\max}\) (mg/L), and liver concentration at 2 hours (mg/L). These represent total exposure, peak systemic concentration, and an early tissue prediction.
For global screening, suppose the input ranges are specified as follows: \(CL_{\mathrm{int}}\), \(Q_h\), and \(f_u\) vary from 0.8 to 1.2 times their reference values; \(k_a\) from 0.75 to 1.25 times its reference value; and \(CL_r\) from 0.5 to 1.5 times its reference value. The global analysis uses independent uniform sampling over these ranges. This independence is an explicit design assumption for this example, not a claim that the biological inputs are independent in all populations. If dependence is relevant, a method appropriate for correlated inputs is needed.
Local and global sensitivity answer different questions. A local elasticity describes a model’s response near one reference parameter vector. Sobol first-order indices describe variance contributions over the specified input domain under the stated sampling assumptions; total-effect indices also include contributions involving interactions. Neither ranking, by itself, establishes that a parameter can safely be fixed.
Worked Example: Calculating a Local Elasticity
At the reference input values, consider the AUC response to \(CL_{\mathrm{int}}\). The hypothetical simulation gives AUC \(107.45\) mg·h/L when \(CL_{\mathrm{int}}\) is set to its reference value divided by \(1.10\), and \(93.6\) mg·h/L when it is multiplied by \(1.10\). These reciprocal perturbations are symmetric on the log-parameter scale.
A central finite-difference estimate of elasticity is
Substitution gives \(\ln(93.6/107.45)\approx -0.13800\) and \(2\ln(1.10)\approx 0.19062\), so \(\widehat{E}_{CL_{\mathrm{int}}}\approx -0.7239\). Near the reference values, a 1% increase in intrinsic clearance is associated in the model with approximately a 0.724% decrease in AUC. This is a local model response, not a causal effect in patients and not a guarantee that the same elasticity holds across the full input range.
Compare Local and Global Evidence
The following hypothetical results summarize the screening. The local elasticities are evaluated at the reference input vector. The first-order and total-effect indices are global estimates over the ranges specified above; small differences between the two columns can reflect interactions as well as simulation estimation error. The indices are not percentages of a common total that must sum to one, especially the total-effect indices.
| Input | AUC local elasticity | AUC first-order index | AUC total-effect index | Screening interpretation |
|---|---|---|---|---|
| \(CL_{\mathrm{int}}\) | -0.724 | 0.50 | 0.61 | Important for AUC; retain |
| \(Q_h\) | -0.17 | 0.07 | 0.13 | Some influence; retain for scenario coverage |
| \(f_u\) | -0.20 | 0.15 | 0.21 | Material influence; retain |
| \(k_a\) | 0.02 | 0.01 | 0.03 | Small for AUC, but relevant to peak output |
| \(CL_r\) | -0.04 | 0.02 | 0.04 | Small in the tested domain; candidate to fix |
The other selected outputs change the interpretation of this table. In the hypothetical analysis, the global total-effect indices for \(C_{\max}\) are 0.29 for \(CL_{\mathrm{int}}\), 0.08 for \(Q_h\), 0.30 for \(f_u\), 0.24 for \(k_a\), and 0.02 for \(CL_r\). For liver concentration at 2 hours, they are 0.38, 0.14, 0.20, 0.23, and 0.02, respectively. Thus \(k_a\) is not a reasonable candidate to fix merely because its AUC sensitivity is low. By contrast, \(CL_r\) is low across all three chosen outputs within the stated range.
Worked Example: Reading the Global Screening Results
For AUC, the estimated first-order index for \(CL_r\) is 0.02 and its total-effect index is 0.04. In this hypothetical analysis, the first-order contribution is about 2% of the output variance over the sampled domain, while the total-effect contribution, including interactions, is about 4%. The difference is small, so there is little evidence that interactions involving \(CL_r\) materially increase its contribution within these ranges.
That result does not alone justify fixing \(CL_r\). The screening must also be checked for adequate numerical precision, evaluated for all selected outputs, and challenged across the intended scenarios. If the scientific use included a drug with substantial renal elimination, a population with altered renal function, or a wider \(CL_r\) range, this same ranking would not settle the decision. The conclusion is limited to this drug-model example and its stated domain.
Reduce the Set and Verify Predictions
A useful reduction rule is set before comparing the full and reduced models. For this example, the team decides that fixing \(CL_r\) is acceptable only if each selected prediction differs from the full model by no more than 5% in every prespecified scenario. This is an illustrative tolerance, not a universal standard; a real analysis should justify its criterion based on the decision and the scale of relevant errors.
The reduced model keeps \(CL_r\) in the equations but fixes it at the reference value, \(0.80\) L/h. The other four parameters remain available to vary. The full-model comparison uses the same parameter values for those four inputs and the scenario-specific physiology; only \(CL_r\) differs when its full-model value is not the reference value.
Worked Example: Checking a Dose Scenario
For the hypothetical high-dose scenario, the full PBPK model predicts AUC \(1260\) mg·h/L, \(C_{\max}=82.0\) mg/L, and liver concentration at 2 hours of \(47.0\) mg/L. With \(CL_r\) fixed at \(0.80\) L/h, the reduced model predicts \(1248\) mg·h/L, \(80.9\) mg/L, and \(46.2\) mg/L, respectively.
For AUC, the relative difference is
For \(C_{\max}\), it is \(100(80.9-82.0)/82.0=-1.341\%\). For the liver concentration, it is \(100(46.2-47.0)/47.0=-1.702\%\). All three absolute differences are below the 5% criterion in this scenario. The negative signs show that the reduced-model predictions are lower than the corresponding full-model predictions; they do not indicate statistical significance.
The remaining scenario checks must be interpreted in the same way. Suppose at low hepatic blood flow and the reference dose, the full-model outputs are AUC \(510\) mg·h/L, \(C_{\max}=29.2\) mg/L, and liver concentration \(22.1\) mg/L. The reduced model gives \(505\) mg·h/L, \(28.7\) mg/L, and \(21.8\) mg/L. The relative differences are \(-0.9804\%\), \(-1.712\%\), and \(-1.357\%\), respectively. These all pass the illustrative criterion. Together with the high-dose check, this supports fixing \(CL_r\) for the tested outputs and scenarios, but does not certify the reduction for untested applications.
Derivation: Why a Log-Scale Difference Estimates Elasticity
For a positive output \(Y\) and positive parameter \(\theta\), elasticity is \(E_\theta=\partial\ln(Y)/\partial\ln(\theta)\). By the chain rule, this equals \((\theta/Y)(\partial Y/\partial\theta)\). It is therefore a dimensionless measure of proportional response: if the model is locally smooth, a small proportional change in \(\theta\) produces approximately \(E_\theta\) times that proportional change in \(Y\).
A central difference on the log scale approximates that derivative. With a multiplicative perturbation \(c>1\), the log-parameter values are \(\ln(\theta c)\) and \(\ln(\theta/c)\), whose difference is \(2\ln(c)\). Subtracting the corresponding log outputs and dividing by this parameter difference gives the estimate used above. The approximation is most informative when the model response is smooth across the perturbation; if the response is highly nonlinear, repeat at smaller and larger perturbations rather than treating one finite difference as a universal sensitivity.
Workflow, Diagnostics, and Alternatives
Name the predictions and scenarios the reduced model must preserve; choose endpoints that reflect its intended use.
Compare local elasticities with global indices over documented parameter ranges. Examine all outputs, not only a convenient endpoint.
State which input is fixed, its fixed value, and the acceptable prediction difference before checking the reduced model.
Compare full and reduced predictions over representative doses and physiological conditions, then document any failures and the limits of the conclusion.
Several diagnostics help identify an unreliable reduction. A parameter may rank low locally but high globally, which suggests that its effect changes away from the reference point. A sizable gap between first-order and total-effect indices indicates interactions worth investigating. Rankings that change sharply with the endpoint, dose, or physiology show that a single pooled ranking is insufficient. Finally, unstable indices across repeated or enlarged simulation designs can mean the sensitivity estimates are too imprecise to support a reduction.
A local analysis is economical and directly describes nearby model behavior, but can miss nonlinear or scenario-dependent effects. A global analysis covers a stated range and can reveal interactions, but its result depends on the range and input-dependence assumptions. Fixing a low-influence parameter is a direct reduction, while lumping or reparameterizing inputs may be preferable when multiple parameters have similar or poorly distinguishable effects. Those alternatives change the model’s parameterization and require their own prediction checks; they are not interchangeable with simply fixing a parameter.
A common mistake is to argue, “The AUC index for \(CL_r\) is small, so renal clearance can be removed.” This reasoning overlooks other outputs and scenarios, treats a domain-specific sensitivity as universal, and confuses fixing an input with deleting its process. The correction is to retain the process at a specified value, examine sensitivity for every decision-relevant output, and test prediction preservation in the scenarios the reduced model is expected to cover.
Check Your Understanding
Answer each question using the definitions and hypothetical workflow above. Short solutions follow the questions.
- Why is a small local elasticity for \(k_a\) with respect to AUC not sufficient evidence to fix \(k_a\) in this example?
- If a reduced-model prediction is 95 mg/L and the full-model prediction is 100 mg/L, calculate the relative difference and decide whether it passes a 5% absolute tolerance.
- What does a gap between a parameter’s first-order and total-effect indices suggest?
- Why should a fixed parameter remain in the PBPK equations rather than be deleted?
- Name two features of the analysis that limit the conclusion that \(CL_r\) can be fixed.
Solutions. (1) \(k_a\) has a larger global total-effect index for \(C_{\max}\), so fixing it could impair a selected prediction even if its AUC effect is small. (2) \(100(95-100)/100=-5\%\); it is exactly at the tolerance boundary and passes an “at most 5%” rule. (3) It suggests that interactions contribute to the parameter’s total effect over the analyzed domain. (4) Fixing retains the modeled process and its contribution at the stated value; deleting it changes the model structure. (5) The conclusion applies only to the specified parameter ranges, outputs, sampling assumptions, dose and physiology scenarios, and error tolerance; any two of these are sufficient examples.