Why One-At-a-Time Sensitivity Can Mislead

What One-at-a-Time Sensitivity Can—and Cannot—Show

A PBPK simulation may depend on many physiological, anatomical, and drug-specific parameters. One-at-a-time (OAT) sensitivity analysis changes one parameter while holding all others at their baseline values, then records how a selected model output changes. It is a useful way to check a model locally and to communicate the effect of a defined perturbation. But an isolated change answers a conditional question: what happens along this particular coordinate direction, with the other parameters fixed? It does not, by itself, establish how the output behaves across the plausible parameter space.

That distinction matters when parameters interact, the response is nonlinear, or multiple parameter changes offset one another. An OAT result can then understate joint effects, fail to represent behavior away from the baseline, or suggest that an endpoint constrains parameters more clearly than it does. The issue is not that OAT calculations are inherently incorrect. It is that their conclusions must be limited to the perturbations, baseline, and endpoint actually examined.

Definition: Let \(Y=f(\theta_1,\ldots,\theta_p)\) be a model output at a specified time or summary endpoint, and let \(\boldsymbol{\theta}_0\) be the baseline parameter vector. An OAT perturbation of parameter \(\theta_j\) evaluates \(f\) at a changed value of \(\theta_j\) while holding every \(\theta_k\), \(k\ne j\), at its baseline value. A finite OAT effect is the resulting output difference or ratio for that specified perturbation. It is conditional on the baseline and perturbation size; it is not a summary of all joint parameter changes.

Perturbations are often easier to interpret on a relative scale. Define \(x_j=\log(\theta_j/\theta_{j,0})\), where \(\theta_{j,0}\) is the baseline value of a positive parameter. Then \(x_j=0\) at baseline, and a given change in \(x_j\) represents a multiplicative change in \(\theta_j\). This scale is not mandatory, but the scale used should be stated: a fixed absolute change and a fixed percentage change need not give comparable results across parameters.

Interactions: Isolated Effects Need Not Add

An interaction means that the effect of changing one parameter depends on the value of another. For two parameters, a practical check is a joint-change contrast. Let \(Y_{00}\) be the baseline output, \(Y_{10}\) the output after changing only parameter 1, \(Y_{01}\) the output after changing only parameter 2, and \(Y_{11}\) the output after changing both. On the output scale being examined, the contrast

\Delta_{\mathrm{int}}=Y_{11}-Y_{10}-Y_{01}+Y_{00}

is zero when the joint change equals the sum of the two isolated changes relative to baseline. A nonzero contrast signals non-additivity for those particular changes on that scale. The contrast is scale-dependent: additivity of \(Y\) is not generally the same as additivity of \(\log Y\).

Justification: For a smooth response \(g(x_1,x_2)\), expand the four values around the baseline \((0,0)\). Subtracting \(g(d,0)\) and \(g(0,d)\) from \(g(d,d)\), then adding \(g(0,0)\), cancels the baseline and the first-order terms. The leading remaining term is the mixed derivative multiplied by the perturbation sizes:

g(d_1,d_2)-g(d_1,0)-g(0,d_2)+g(0,0) = \frac{\partial^2 g}{\partial x_1\partial x_2}(0,0)d_1d_2 +\text{higher-order terms}.

Thus, for small perturbations, the contrast reflects local interaction curvature; for larger perturbations, it also includes higher-order behavior across the selected rectangle. A zero contrast at one pair of perturbations does not prove that interactions are absent elsewhere.

Worked Example: A Joint Change Adds More Than the Isolated Effects

Consider a hypothetical PBPK output \(Y\), expressed as a ratio to its baseline value. Suppose its local response surface in two log-scaled parameter coordinates is

Y=\exp(0.2x_1+0.2x_2+0.8x_1x_2).

The interaction term is part of this illustrative response surface; the equation is not proposed as a universal PBPK mechanism. At baseline, \(x_1=x_2=0\), so \(Y_{00}=1\). Perturb each parameter separately to \(x=0.5\), holding the other at zero. The two OAT results are identical:

Y_{10}=Y_{01}=\exp(0.2\times0.5)=\exp(0.1)=1.105.

If the effects were additive on the log-output scale, the joint prediction would be \(\exp(0.1+0.1)=\exp(0.2)=1.221\). But when both parameters change, the interaction term contributes \(0.8\times0.5\times0.5=0.2\) to the log output:

Y_{11}=\exp(0.1+0.1+0.2)=\exp(0.4)=1.492.

The interaction contrast on the log scale is \(0.4-0.1-0.1+0=0.2\), not zero. The joint output is 1.492 times baseline, whereas adding the two isolated log effects would predict 1.221 times baseline. OAT accurately describes each isolated perturbation, but does not reveal the extra joint effect unless a joint change is also evaluated.

Nonlinearity: A Local Perturbation May Not Represent a Range

Even without an interaction between two parameters, the response to one parameter can change with its own value. A small perturbation near a flat part of a response curve may produce a small OAT effect, even though the output changes sharply elsewhere. Conversely, a perturbation near a steep transition may exaggerate the effect expected over most of the plausible range. A local derivative or finite difference describes behavior near the point and at the scale assessed; it is not a guarantee of a constant effect over a wider range.

Worked Example: A Small Local Change Misses a Response Transition

Suppose, for illustration, that a dimensionless PBPK endpoint \(R\), interpreted as a fraction between 0 and 1, follows a steep saturating response to a positive parameter ratio \(z\):

R(z)=\frac{z^8}{1+z^8}.

At the baseline \(z=0.50\), \(z^8=0.003906\), giving \(R=0.003891\), or 0.3891%. A 10% downward perturbation gives \(z=0.45\), \(z^8=0.001682\), and \(R=0.001679\), or 0.1679%. A 10% upward perturbation gives \(z=0.55\), \(z^8=0.008373\), and \(R=0.008304\), or 0.8304%. These small OAT changes are consistent with a response that is still near its lower plateau.

They do not represent the whole curve. At \(z=1\), \(R=0.5\), or 50.0%; at \(z=1.5\), \(z^8=25.63\) and \(R=0.9624\), or 96.24%. Thus, if values from 0.45 to 1.5 are scientifically plausible, inspecting only a 10% perturbation around 0.50 would miss the large transition. This example diagnoses a general risk; the equation is a hypothetical response, not a claim that every PBPK endpoint has this form.

Perturbation size is therefore part of the result, not a minor implementation detail. Report the baseline, direction and magnitude of each change, the output scale, and whether the change is absolute, percentage-based, or defined on a transformed parameter scale. Where plausible ranges are broad, evaluate more than one perturbation size or inspect the response across a range. A one-sided perturbation can also conceal asymmetry: the response to increasing a parameter may differ from the response to decreasing it by the same percentage.

Parameter Compensation: One Endpoint Can Stay the Same

Two or more parameters may change in opposite directions so that one endpoint is preserved. This is parameter compensation. It is especially important when interpreting calibration: a good fit to one summary endpoint does not necessarily mean that the individual parameters are uniquely determined. An OAT analysis can show that each parameter affects the endpoint when varied alone, while failing to show that a coordinated change in both leaves it unchanged.

Worked Example: Bioavailability and Clearance Compensate for AUC

For a hypothetical linear oral model, suppose the AUC is \(AUC=Dose\times F/CL\). The administered dose is 100 mg, baseline bioavailability is \(F=0.50\), and clearance is \(CL=10\) L/h. Then

AUC=\frac{100\ \mathrm{mg}\times0.50}{10\ \mathrm{L/h}} =5.00\ \mathrm{mg\cdot h/L}.

Increasing \(F\) by 20% gives \(F=0.60\), with clearance fixed, and \(AUC=100\times0.60/10=6.00\ \mathrm{mg\cdot h/L}\). Increasing \(CL\) by 20% gives \(CL=12\) L/h, with bioavailability fixed, and \(AUC=100\times0.50/12=4.167\ \mathrm{mg\cdot h/L}\). Each OAT change has a clear effect.

If both parameters increase by 20%, however, \(AUC=100\times0.60/12=5.00\ \mathrm{mg\cdot h/L}\), exactly the baseline value. The equality follows because \(F/CL\) is unchanged: \(0.60/12=0.50/10\). This compensation concerns AUC only. It does not establish that the concentration-time profiles, peak concentrations, or other model outputs are equal under the two parameter sets.

Compensation is endpoint-specific. AUC can remain constant while absorption rate, tissue distribution, or the timing of concentrations differs. A robust assessment therefore examines outputs relevant to the scientific or regulatory question, rather than assuming that agreement for one endpoint guarantees agreement elsewhere. This is also a reason to inspect simulated time courses when the timing of exposure matters.

Practical Checks and Alternatives

OAT analysis is most informative when the baseline is clearly defined, the perturbations have a scientific rationale, and the output and scale are specified. Interpret its results as conditional changes around that baseline. Do not infer that a low OAT effect makes a parameter unimportant under all conditions, or that a high OAT effect means it is independently estimable from data.

Useful diagnostics include comparing upward and downward perturbations, repeating perturbations at different magnitudes, checking selected pairs jointly, and examining more than one endpoint. For pairwise checks, the four-run contrast above distinguishes the joint result from the sum of isolated effects on the chosen scale. If the model has thresholds, saturating processes, or parameters that plausibly vary together, extend the checks beyond a single baseline point. Large discontinuities or unexpected directional changes also warrant checking numerical convergence and model implementation before interpreting them as biological behavior.

The main alternatives answer different questions. Local derivative and second-order analyses characterize slopes and curvature near a specified parameter set. Scenario analyses compare scientifically defined combinations of parameter values. Broader joint-variation approaches assess behavior over a parameter region, but require choices about ranges or distributions and can be more computationally demanding. None removes the need to select relevant outputs and interpret results in context.

Approach	What it can show	What it does not establish alone
One-at-a-time perturbation	Effect of a specified isolated change at a chosen baseline	Joint effects, behavior across broad ranges, or compensation
Joint pair perturbation	Whether two selected changes combine additively on a specified output scale	All interactions among all parameters or across the entire parameter space
Local derivative or curvature	Local slope or curvature near a specified parameter set	Global behavior far from that set
Scenario comparison	Consequences of prespecified combinations relevant to a question	Effects of combinations not included in the scenarios

A common mistake is to rank parameters by their OAT effects and conclude that the largest-ranked parameter is always the most influential, or that the lowest-ranked parameter can safely be ignored. The correction is to state the ranking narrowly: it orders effects for the chosen baseline, perturbation, output, and scale. Before using it to guide a decision, check whether plausible joint changes, nonlinear response regions, or endpoint-specific compensation could change the conclusion.

Key takeaway: OAT sensitivity is a conditional probe, not a complete map of PBPK model behavior. Use joint contrasts to check interaction, multiple perturbation sizes to check nonlinearity, and multiple relevant outputs to detect compensation that preserves only one endpoint.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

If the isolated changes in a log-transformed output are 0.10 and 0.15, what log-output change would additivity predict for the joint change? What would a measured joint change of 0.35 indicate?
Why does a small OAT effect near the baseline not demonstrate that a parameter has little influence over its full plausible range?
In the AUC example, why does increasing both \(F\) and \(CL\) by 20% preserve AUC? Does that prove the concentration-time profiles are identical?
What four model evaluations are needed to compute a two-parameter interaction contrast relative to a baseline?

Solutions. (1) Additivity predicts \(0.10+0.15=0.25\). A joint change of 0.35 gives a contrast of \(0.35-0.10-0.15=0.10\), indicating non-additivity on the log-output scale for those perturbations. (2) The response may be nonlinear: the local slope near baseline can be small even when the output changes sharply elsewhere. (3) The ratio \(F/CL\) is unchanged because both numerator and denominator are multiplied by 1.20. This preserves the specified AUC formula, but does not prove that the time courses or other endpoints are identical. (4) Evaluate the baseline, each of the two single-parameter changes, and the joint change; combine them as \(Y_{11}-Y_{10}-Y_{01}+Y_{00}\) on the chosen output scale.