Why Parameter Reduction Must Be Sequential
A PBPK model may contain many parameters, but not every parameter matters equally for every prediction. The relevant question is not simply which parameter has the smallest sensitivity ranking. It is which parameters, or combinations of parameters, are needed to represent a specified set of outputs over a specified domain. A parameter can have little influence on one endpoint and substantial influence on another, or matter only at a particular dose, time, or physiological scenario.
A defensible reduction therefore has stages. First define what the model must preserve. Then use screening to find candidates for closer study, examine those candidates using endpoint-appropriate sensitivity and identifiability information, and only then decide whether to retain, fix, group, or reparameterize them. The earlier material on fixing low-influence parameters, lumping parameters with similar effects, and reparameterizing informative parameter combinations supplies the individual tools; the focus here is how to order and connect them.
The workflow assumes that the model structure, target outputs, plausible parameter ranges, and relevant conditions have been specified. The simulation implementation must also be reliable enough that apparent output changes reflect parameter changes rather than solver failure or numerical noise. If a local derivative is used, the output must be sufficiently smooth near the point where it is evaluated. For screening methods that sample across a domain, the domain and any parameter dependencies must be represented appropriately.
Stage 1: Define the Reduction Target
Before ranking parameters, write down the outputs and conditions the reduced model is expected to support. These might include an exposure summary and a concentration at a specified time, for several doses or physiological scenarios. Define the allowable error for each output in units or as a relative error. A single reduction can be adequate for one target and inadequate for another.
Also state the parameter domain over which the decision is intended to hold. A parameter judged negligible in a narrow neighborhood of a nominal value may matter over a wide plausible range. If inputs are dependent, the screening domain should not include combinations that are physiologically or scientifically impossible. Sensitivity results are conditional on these choices; they do not establish universal parameter importance.
Stage 2: Screen Broadly, Then Investigate Selectively
Screening is a triage step. Morris screening can flag parameters with substantial or variable effects over a sampled range; Sobol indices can describe contributions to output variance under specified sampling assumptions. The earlier material on building a PBPK screening analysis and designing and checking Sobol simulations gives the relevant design considerations. Screening helps decide what to investigate in detail, but a low screening score alone does not prove that a parameter can be removed.
A simple illustrative score makes the decision logic clear. For parameter \(i\), define \(r_{ij}\) as the largest absolute relative change observed in target output \(j\) in a screening design, and define \(R_i=\max_j r_{ij}\). This maximum protects against discarding a parameter that matters to even one of the selected outputs. It is a practical summary for this example, not a universal screening statistic. The threshold for flagging a parameter should be chosen in relation to output tolerances and the purpose of the analysis.
Worked Example: Screening Against Multiple Outputs
Suppose a hypothetical screening exercise varies four parameters over prespecified plausible ranges. It records the largest absolute relative changes in AUC and \(C_{\max}\), relative to their baseline values of \(100\ \mathrm{mg\cdot h/L}\) and \(10\ \mathrm{mg/L}\). The resulting percentages are hypothetical screening summaries, not study findings.
| Parameter | Largest AUC change | Largest \(C_{\max}\) change | \(R_i\) |
|---|---|---|---|
| Absorption rate \(k_a\) | 1% | 18% | 18% |
| Hepatic blood flow \(Q_{\mathrm{hep}}\) | 12% | 10% | 12% |
| Tissue partition parameter \(K_p\) | 2% | 3% | 3% |
| Renal clearance parameter \(CL_{\mathrm{ren}}\) | 0.5% | 0.4% | 0.5% |
If the team uses 3% as a provisional screening flag, then \(k_a\), \(Q_{\mathrm{hep}}\), and \(K_p\) proceed to detailed analysis because their maximum scores are at least 3%. \(CL_{\mathrm{ren}}\) is a candidate for reduction, not an automatic deletion: its effects must still be interpreted against the intended domain, output tolerances, and model context. The endpoint maximum matters here. Ranking by AUC alone would make \(k_a\) appear unimportant even though its screened effect on \(C_{\max}\) is 18%.
Stage 3: Examine Candidate Parameters in Detail
For parameters that pass screening, use more detailed sensitivity information at relevant outputs and conditions. Local derivatives can show direction and magnitude near a reference parameter vector; global analyses can reveal effects over a broader domain. The earlier material on time- and endpoint-dependent rankings, second-order sensitivity, and global sensitivity describes why neither one-at-a-time results nor one ranking should be treated as a complete account.
For positive parameters and outputs, a useful local summary is the elasticity \(S_{ji}=\partial\ln g_j/\partial\ln\theta_i\). Near a reference point, a small proportional change in parameter \(i\) gives an approximate proportional output change \(S_{ji}\,\Delta\ln\theta_i\). This is a local approximation, not a guarantee over a wide range. A first-order bound can help prioritize investigation: if plausible changes satisfy \(|\Delta\ln\theta_i|\leq \delta_i\), then the sum \(\sum_i |S_{ji}|\delta_i\) bounds the magnitude of the linearized change in output \(j\). It does not bound nonlinear behavior unless additional conditions justify that use.
Worked Example: Comparing Local Effects With a Tolerance
At a hypothetical reference parameter vector, suppose the elasticities for AUC and \(C_{\max}\) with respect to \(k_a\), \(Q_{\mathrm{hep}}\), and \(K_p\) are:
Suppose plausible one-sided changes from the reference are 10% for \(k_a\), 15% for \(Q_{\mathrm{hep}}\), and 30% for \(K_p\). Their log-scale changes are \(\ln(1.10)=0.0953\), \(\ln(1.15)=0.1398\), and \(\ln(1.30)=0.2624\). For \(K_p\), the largest first-order relative change across the two outputs is approximately \(0.10(0.2624)=0.0262\), or 2.62% for AUC; for \(C_{\max}\), it is \(0.05(0.2624)=0.0131\), or 1.31%. If the stated tolerance is 5% for both outputs, this local calculation supports treating \(K_p\) as a fixing candidate near this reference point. It does not establish that fixing it is safe across a wider domain or in the presence of important interactions.
For comparison, the \(k_a\) contribution to \(C_{\max}\) is approximately \(0.85(0.0953)=0.0810\), or 8.10%, which exceeds the tolerance. The \(Q_{\mathrm{hep}}\) contribution to AUC is approximately \(0.60(0.1398)=0.0839\), or 8.39% in magnitude. These parameters should remain represented for these targets under the stated ranges and tolerance. The negative elasticity means the corresponding output decreases as the parameter increases locally; the magnitude is what is compared with a symmetric error tolerance.
Stage 4: Choose a Reduced Representation
After detailed analysis, make an explicit choice for each parameter or parameter combination. Retain parameters with material effects on any target. Fix a parameter only when its influence is acceptably small for the intended domain, or when external information justifies holding it at a specified value. Group or reparameterize parameters when the outputs depend on an informative combination and the discarded directions are not needed for the chosen targets. The earlier material on sensitivity-based identifiability diagnostics is relevant: weakly distinguishable parameters do not necessarily imply that their individual effects are negligible.
When fixing a parameter, record its fixed value and why that value is defensible. When replacing several parameters with a combination, state the transformation and what information is lost. Similar sensitivity columns may motivate investigating a combination, but similarity at one point is not proof that the same combination represents the model across the domain. If the PBPK equations show that relevant outputs depend exactly on a product, using that product can be a structural reduction; otherwise, it is an approximation whose scope must be stated.
Worked Example: Replacing Two Parameters With a Product
Consider a hypothetical transport component in which the selected outputs depend on permeability \(P\) and membrane area \(A\) only through the product \(\psi=PA\). At the reference values \(P=0.40\) in its specified units and \(A=2.00\) in its specified units, \(\psi=0.40\times2.00=0.800\) in product units. The model equations for the chosen outputs contain \(\psi\), but do not contain \(P\) or \(A\) separately.
For those outputs, replacing \(P\) and \(A\) by \(\psi\) preserves the modeled transport term exactly. For example, \(P=0.50\) and \(A=1.60\) also give \(\psi=0.50\times1.60=0.800\). The two parameter pairs therefore give the same value of this term, and hence the same selected outputs under the stated structural assumption. The reduced representation has one parameter rather than two.
This conclusion does not extend automatically to other outputs or to a model in which \(P\) and \(A\) also enter separately. Before using the product representation, inspect the equations and the chosen outputs. If either parameter has another pathway into a target, or if the outputs must distinguish their individual values, the reduction is not exact for that purpose.
Order, Diagnostics, and Common Failure Modes
The stages are sequential because each answers a different question: screening asks what merits attention; detailed sensitivity asks how candidate parameters affect chosen outputs; reduction asks what simpler representation is justified. Revisit earlier decisions if detailed analysis exposes a missed endpoint, a nonlinear response, or an important interaction. Keep a record of the domain, reference values, output tolerances, sensitivity method, and rationale for each fix or combination so the reduced parameterization has an interpretable scope.
Warning signs include rankings that change sharply across time points or scenarios, local derivatives that vary greatly across the parameter range, screening results that depend strongly on the design, and a reduced representation that cannot express a scientifically plausible direction of change. Also investigate solver warnings, failed runs, and unexpected discontinuities before interpreting sensitivity scores. A numerical artifact can look like an influential effect, while failed simulations can hide one.
The main alternatives are not interchangeable. Screening methods offer broad coverage at relatively low cost but provide limited detail about a particular target or local direction. Local sensitivity analysis is efficient and interpretable near a reference point, but can miss nonlinearities and interactions over a wider domain. Detailed global analyses examine broader ranges but require more computation and depend on the assumed sampling domain and parameter dependence. Fixing reduces dimension directly but discards variation in the fixed parameter; grouping can preserve an influential combination while losing separate parameter interpretation. The appropriate sequence uses screening to focus effort, then chooses among these reduction options according to the outputs and assumptions.
Parameter reduction is always purpose-specific. It does not establish that the removed parameters are biologically unimportant, nor does a successful reduction for one prediction set guarantee suitability for another. The next step in a modeling workflow is to assess how the chosen reduced representation behaves for its intended predictions; the decisions made here should leave that assessment a clear record of what was reduced and why.
Name outputs, conditions, plausible parameter domain, and acceptable output errors.
Use a design suited to the domain and flag candidates without treating low ranks as automatic permission to delete.
Examine endpoint-specific, time-dependent, and where needed interaction or global sensitivity evidence.
Retain, fix, group, or reparameterize with an explicit rationale, fixed values, and scope.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- Why should a screening rank not be used by itself to decide that a parameter is irrelevant?
- A parameter has elasticities 0.02 for AUC and 0.70 for \(C_{\max}\). Which output should guide a reduction intended to preserve both?
- For an elasticity of \(-0.40\) and a plausible 10% parameter increase, calculate the approximate local proportional output change using \(\ln(1.10)=0.0953\).
- Under what condition can two parameters be replaced by their product without changing selected model outputs?
Solutions. (1) The rank depends on the sampled domain, design, and selected outputs, and may not reflect nonlinearities or interactions; it identifies candidates for further analysis. (2) Both outputs matter, so the larger \(C_{\max}\) effect must be considered rather than relying on AUC alone. (3) The approximation is \(-0.40(0.0953)=-0.0381\), or a 3.81% decrease locally. (4) The selected outputs must depend on the parameters only through their product over the scope being claimed; separate pathways or other targets can invalidate the reduction.