Matching Model Detail to the Prediction Target

Choose Detail for the Prediction You Need

A PBPK model can represent tissues as individual organs or combine several tissues into a smaller number of pooled compartments. Neither choice is automatically better. The relevant question is whether the chosen structure can provide the prediction needed for a particular decision. A model intended to predict plasma exposure may not need to report every organ concentration. A model intended to predict concentration in a named organ must, however, preserve enough information to distinguish that organ from the tissues pooled with it.

The target should be stated before comparing structures: specify the quantity, its location or aggregation, the time period, and any acceptable prediction error. “Predict tissue distribution” is too vague. “Predict the volume-weighted mean concentration across liver, muscle, and adipose tissue at 24 hours” is a defined target, as is “predict kidney concentration at 24 hours.”

Three common targets illustrate why the choice matters. Plasma concentration or plasma area under the concentration–time curve (AUC) is a systemic target. Average tissue burden combines information across a defined set of tissues. Concentration in a particular organ is a local target. These outputs are related, but one cannot generally be substituted for another: a correct plasma profile does not by itself establish correct tissue concentrations.

Define the Aggregation Before Pooling

Suppose a target averages concentrations across \(m\) tissues. Let \(V_i\) be the volume of tissue \(i\), and let \(C_i(t)\) be its concentration at time \(t\). If the target is the mean concentration over the combined tissue volume, its value is

The weights are tissue volumes because this definition averages concentration across the combined volume. The total amount in those tissues is a different quantity, \(\sum_i V_i C_i(t)\), with units of amount rather than concentration. An unweighted mean, \(\sum_i C_i(t)/m\), answers yet another question: it gives every tissue the same weight regardless of size. The target definition, not convenience, determines which summary is appropriate.

The same weighting applies to AUC when the set of tissues and their volumes are fixed. Integrating the volume-weighted concentration over time gives

This follows because the volumes are constant: the finite weighted sum can be taken outside the integral. It does not imply that a single pooled compartment will reproduce each tissue’s trajectory. It establishes how to calculate the specified aggregate from organ-resolved predictions.

Match Structure to Three Common Targets

For a plasma target, a lumped tissue representation may be sufficient if it reproduces plasma predictions accurately across the relevant dose, time range, and conditions. Plasma data alone do not establish that tissue predictions are accurate: different tissue distributions can be compatible with similar plasma profiles. If the application only uses plasma exposure, that limitation may be acceptable. If tissue concentrations affect the decision, plasma agreement is not enough.

For an average tissue-burden target, pooling can be appropriate when the required output is specifically the defined average, and the pooled representation preserves that average over the relevant conditions. It may be unnecessary to estimate separate organ concentrations when no decision depends on them. But a pooled state is not automatically equal to a volume-weighted mean: the model’s pooled quantity and the target’s weights must actually correspond.

For a named-organ target, the model must distinguish that organ or provide an independently justified mapping to its concentration. A pooled prediction can sometimes approximate one organ under restricted conditions, but it does not generally identify that organ’s value. When organs differ in flow, tissue partitioning, or kinetic behavior, a shared state can conceal differences that matter to the target.

Justify the Choice and Check for Failure

A useful selection procedure is to make the target explicit, test whether the simpler structure preserves it, and add resolution only when that structure cannot meet the target’s requirements. This is consistent with the earlier material on model structure as a prediction choice, pooling tissues, and structural error: the comparison must focus on the output needed, rather than on compartment count alone.

The comparison is only informative when the predictions are evaluated on the same target, conditions, and benchmark. A model that is calibrated and assessed on the same observations may appear accurate without demonstrating performance for new predictions. Check the time course, not just one summary value, when the intended target is an exposure profile or AUC. For average burden, verify that the tissue set, volumes, and weighting in the calculation match the target definition.

Warning signs include a pooled concentration being labelled as if it were an organ-specific measurement; a model matching plasma while missing tissue observations; or an average that changes materially when the weights are corrected. Large, systematic discrepancies for one organ suggest that the pooled structure may be hiding relevant heterogeneity. Conversely, if organ-resolved outputs are not supported by relevant data or do not improve the target prediction, their apparent detail should not be mistaken for established accuracy.

Common Mistake and Practical Limits

Wrong reasoning: “The organ-resolved model has more compartments, so it must be the better model for every prediction.” More detail can provide outputs unavailable from a pooled model, but it does not guarantee that those outputs are accurate. Correction: judge both structures against the stated prediction target and its required accuracy. If only plasma AUC is needed, organ-specific states may add complexity without improving the relevant prediction. If kidney concentration is needed, plasma fit or a whole-body average cannot establish that kidney concentration is correct.

This approach has limits. A target-specific comparison depends on having suitable observations or a defensible benchmark, and a tolerance appropriate to the intended application. Results may not carry over to different doses or physiological conditions. A structure adequate for one target may fail for another: a model that predicts plasma exposure adequately may not predict an organ concentration, while a model that captures a named organ may not automatically provide a correct average over a differently defined tissue set. State the scope of the conclusion rather than calling one structure universally adequate.