Model Structure as a Prediction Choice

Why Structure Is a Prediction Choice

A PBPK model is not just a list of physiological values. It also specifies what parts of the body are represented separately, how they are connected, and which processes move or remove drug between them. Those choices determine what the model can distinguish in its predictions. A model with a single shared tissue state, for instance, cannot report a different concentration for each tissue represented by that state, no matter how its parameter values are tuned.

This makes structure a decision about the question to be answered. A model intended to predict a whole-body or plasma concentration may not need to distinguish every tissue. A model intended to predict concentrations in particular organs must represent those organs in a way that permits their predictions to differ. More detail is not automatically better: it also brings additional assumptions and quantities that may be difficult to specify or evaluate.

States, Connections, and Predictions

A compartment is a region that the model treats as a state-bearing unit. Its state may be the amount of drug in that region; if its volume is fixed, the model can express concentration as amount divided by volume. A connection specifies a route by which drug can move between represented regions. Exchange processes describe how drug crosses that route, while removal processes describe drug leaving the modeled system or being eliminated within it.

A concise way to describe a dynamic model is to write its states and outputs as functions of time. Let \(S\) denote the chosen structure, \(x(t)\) the vector of model states, \(u(t)\) the administered input, and \(\theta\) the numerical parameters. Then a model can be represented abstractly as

Here \(f\) describes how the states change, and \(g\) describes which predictions are reported. The structure \(S\) affects both: it defines which states exist and which processes connect them, and it constrains how a requested output can be calculated. This notation is general; a particular PBPK model must still specify its own states, inputs, processes, and observation outputs.

Any numerical illustration of a state equation requires explicit assumptions. For example, to infer a concentration change from an amount change, the relevant volume must be fixed. To attribute an amount change to one listed exchange, one must also account for every other input, output, and loss. A statement such as “there is one input” does not establish that there are no drug losses or other outflows.

How the Structure Restricts What Can Be Predicted

The restriction in the example follows directly from the model’s output definition. If two reported outputs are both defined as the same state, they are equal at every time at which that state is defined. Parameters can affect the value of the shared state, but equality remains an identity of the model. To permit unequal predictions, the model needs separate states or another explicitly defined output mechanism that can yield different values.

The same reasoning applies to connections and processes. If a proposed route for drug exchange is absent, the model cannot represent drug transfer along that route as a distinct process. If a removal process is excluded, it cannot contribute to the modeled rate of change. These absences may be reasonable simplifications for a particular prediction target, but they are assumptions—not facts established by a good numerical fit elsewhere.

The previous material on organ blood-flow allocation supplies physiological flow information that can inform model inputs and connections. Flow values alone do not specify the full structure: one must still decide which regions have distinct states and what drug exchange or removal processes are represented. A structurally coherent model must also obey the physiological and bookkeeping constraints relevant to its chosen representation.

Choose Structure by Starting With the Target

A practical comparison begins by stating the prediction target precisely: which concentration or summary, in which region, and over what time range? Then ask whether each candidate structure has states and outputs capable of representing that target. This is a capability check, not a guarantee that the model will predict accurately. A model may be capable of producing a prediction and still have unsuitable parameter values, incorrect assumptions, or inadequate data support.

For a target that is a whole-body summary, a coarser representation may be sufficient if it preserves the information needed for that summary. For targets requiring tissue-specific differences, a representation that forces tissues to share a state cannot answer the question. The main alternative is not “simple versus good” but a choice between a less differentiated representation and one that distinguishes more regions or processes. More differentiated structure can answer more specific questions, but it also requires more assumptions and numerical inputs.

A useful comparison evaluates candidate structures against the same target and the same hypothetical or observed data. The numerical score can summarize discrepancy, but it does not establish that the preferred structure is uniquely correct. A score may favor a model for the selected target and time points while leaving other predictions untested.

Assumptions, Diagnostics, and Limitations

A structure is always a representation rather than a complete replica of the body. Its compartments simplify regions into states; its connections and exchange processes simplify routes of drug movement. The model’s predictions are conditional on those choices, as well as on the numerical values and inputs used with them. In many compartment formulations, each compartment is assumed to be well mixed, but that is a modeling assumption to state and assess—not a property guaranteed by calling a region a compartment.

Basic checks can reveal when a chosen structure or its implementation is going wrong:

• Check the target against the states. If the requested output requires a difference the model mathematically forces to zero, the structure cannot provide that prediction.
• Check the represented processes. Confirm that every process necessary for the intended prediction is included, and that omitted processes are defensible for the question being asked.
• Check consistency. Verify that the stated connections and processes agree with the intended physiological scenario and that amounts are not created or lost unintentionally by the equations.
• Compare predictions with data in context. Persistent discrepancy can indicate an inadequate structure, but it can also arise from uncertain parameters, inputs, measurements, or other assumptions. Discrepancy alone does not identify its cause.
• Check whether extra detail is supported. If a more detailed structure adds states or processes for which the available information offers no meaningful constraint, its outputs may depend strongly on unsupported assumptions.

A common mistake is to treat a poor prediction as proof that one parameter has the wrong value, then keep adjusting parameters even when the structure forces the wrong kind of output. The correction is to check what the model can represent before tuning its values. Conversely, adding compartments merely because they are available is not a reliable remedy: additional states cannot guarantee more accurate predictions when their assumptions or inputs are weak.

Structure comparisons also have limits. Different structures may produce similar predictions for one target while disagreeing for another, and a small set of observations may not distinguish them. A model’s suitability is therefore relative to a stated prediction task and the evidence available for that task. The aim is not maximum detail; it is a representation capable of answering the question with assumptions that can be defended.