Recognizing Unnecessary Structural Complexity

Why Recognize Unnecessary Structure?

A PBPK model can distinguish organs or tissues for sound physiological reasons. But a compartment is not justified merely because an organ exists, or because a parameter can be assigned to it. The practical question is whether the available evidence and the prediction task support that distinction. If observations cannot inform tissue-specific parameters, a more detailed structure may give an appearance of precision without better-supported predictions.

This is a question of model adequacy for a specified use, not a general claim that fewer compartments are always preferable. A pooled model may be appropriate for predicting a plasma profile yet inadequate for a question about concentrations in a particular tissue. The introductory material on model structure as a prediction choice established why the target matters; here the focus is on recognizing signs that a proposed level of detail is not earning its place.

A useful assessment starts with a fair comparison. The candidate models should represent the same dosing scenario, prediction target, and observation process. Their parameters should be estimated using the same data and compatible assumptions about residual error. If one model is given different data, a different target, or a more favorable error treatment, the comparison cannot isolate the consequence of its structure.

What Evidence Would Support the Detail?

A tissue-resolved structure is more defensible when relevant observations distinguish the tissues or when physiological information constrains the parameters that govern their behavior. Tissue concentrations, for example, may directly inform tissue-specific distribution. Plasma observations can also inform tissue parameters indirectly when those parameters have distinct, detectable effects on the plasma predictions. Whether they do so depends on the model, the sampling times, and the variability of the measurements.

By contrast, when only a sparse plasma profile is observed, multiple combinations of tissue-specific parameters may explain the data similarly. A model can still produce numerical estimates, but an optimizer returning values is not proof that the values are well informed. Look for signs such as large changes in a tissue parameter producing negligible changes in observed predictions, or coordinated changes in several parameters leaving those predictions nearly unchanged. Estimates that repeatedly hit imposed bounds are another warning: the data may not locate the parameter within its allowed range.

A direct diagnostic is to perturb a parameter and recalculate the predictions for the observations and target of interest. If a substantial parameter change barely changes those predictions, the data or target provide little leverage on that parameter. Conversely, a clear response does not by itself establish that the parameter is identified: other parameters may produce a similar response. This diagnostic is a way to find questions for further assessment, not a substitute for considering measurement error, physiology, and the intended use.

Compare Structure Against the Prediction Target

The relevant comparison is not “Which model has more compartments?” but “Does the added distinction matter for the prediction I need, and is that distinction supported?” The prediction target should be stated precisely: for example, plasma concentration at specified times, a tissue concentration, or an exposure summary over a stated interval. Then compare candidate structures on that target using observations appropriate to it.

For observed concentration data, one simple descriptive measure is the root mean squared error (RMSE), when the residuals are on a common scale and the observations are being compared consistently. Define residual \(e_i\) as observed concentration minus predicted concentration at observation \(i\). For \(n\) observations, the RMSE is the square root of the average squared residual. It summarizes discrepancy in the same units as concentration; it does not, by itself, establish that a model is physiologically correct or that a difference is important.

For a target that is not directly observed, compare the models' predictions over the relevant scenario and ask whether the difference is large enough to matter for the scientific or decision context. Where possible, use an independent data set or measurements that directly address that target. Do not treat agreement with the data used to fit the model as equivalent to independent validation.

A Short Justification for Sensitivity Checking

The reason parameter perturbation can reveal weak constraint follows from a first-order approximation. Let \(\boldsymbol{\theta}\) be a vector of model parameters, and let \(\mathbf{f}(\boldsymbol{\theta})\) be the vector of predictions at the observed times. For a small parameter change \(\Delta\boldsymbol{\theta}\), differentiability gives \(\mathbf{f}(\boldsymbol{\theta}+\Delta\boldsymbol{\theta})\approx \mathbf{f}(\boldsymbol{\theta})+J\Delta\boldsymbol{\theta}\), where \(J\) is the matrix of prediction sensitivities, with entries \(J_{ij}=\partial f_i/\partial\theta_j\).

If a parameter change has little effect on the predictions, the relevant sensitivity is small. If a nonzero change satisfies \(J\Delta\boldsymbol{\theta}\approx\mathbf{0}\), different parameter combinations produce nearly indistinguishable predictions near the fitted values. This explains why a detailed model can have weakly constrained parameters even when its predictions fit the observed data. The approximation is local: it diagnoses behavior near the current parameter values and does not prove that predictions are identical everywhere.

Alternatives, Diagnostics, and Limits

Pooling similar tissues is one alternative when the target and evidence do not require separate representation; earlier material on pooling tissues with similar kinetics explains conditions under which predictions can be preserved. Keeping the resolved structure is more appropriate when distinct tissue behavior matters and can be supported. A third option is to collect informative observations, such as tissue concentrations or measurements at times when candidate structures predict different behavior, rather than forcing a choice from uninformative data.

Useful warning signs include parameters that change greatly across plausible fits while target predictions remain stable, parameter pairs that compensate for one another, estimates repeatedly reaching bounds, and added compartments that leave the target essentially unchanged. These signs do not automatically prove a compartment should be removed. They prompt a closer look at data quality, parameter constraints, numerical fitting, and whether the chosen target can reveal the proposed distinction.

A common mistake is to reason, “The detailed model fits the observed plasma data better, so its tissue parameters must be correct.” The correction is that improved fit on those observations supports only a narrower claim: the detailed model can reproduce those data more closely under the fitting setup. It does not establish that its added parameters are individually constrained, that the improvement matters for the target, or that tissue predictions are reliable. Likewise, a lack of improvement on one target is not a reason to discard detail needed for another target.

Any judgment is conditional on the evidence, physiological assumptions, measurement error, and intended prediction. A simple model can omit a meaningful process; a detailed one can be hard to constrain. The aim is not minimum complexity at any cost, but a structure whose distinctions are supported and useful for the stated prediction.