Choosing a Model Is a Prediction Decision
A liver model is not just an equation for hepatic clearance. It encodes a claim about how blood and drug mix while passing through the liver. That claim affects predicted extraction and clearance, especially when unbound intrinsic clearance is not small relative to hepatic blood flow. The practical question is therefore not simply which model gives the largest or most familiar estimate, but which model’s assumptions are defensible for the prediction task and which its available data can support.
The well-stirred and parallel-tube formulations represent different limiting pictures of intrahepatic mixing. A dispersion formulation represents axial mixing explicitly. Choosing among them requires considering the intended output, the physiological evidence, the available measurements, and the cost of introducing additional parameters. When evidence does not distinguish the formulations, that uncertainty should be reported rather than hidden by selecting a model solely because it is more complex.
What the Three Formulations Assume
The well-stirred model treats the liver as a single, well-mixed compartment. Drug entering the organ is assumed to mix rapidly, so that the concentration relevant to elimination is the same throughout the compartment. Under the standard steady-state mass-balance assumptions, its extraction ratio is \(E_h=x/(1+x)\), and its clearance is \(CL_h=Q_hx/(1+x)\). It is often a useful baseline when no data justify a more detailed representation of intrahepatic mixing.
The parallel-tube model treats the liver as parallel flow paths with little axial back-mixing. Drug concentration falls along each path as it is exposed to elimination. If paths have the same effective properties, its extraction ratio is \(E_h=1-\exp(-x)\), and its clearance is \(CL_h=Q_h[1-\exp(-x)]\). Its predictions depend on the assumed exposure along the paths, not on an organ-wide concentration that is instantly uniform.
The dispersion model represents concentration changes along the direction of flow while allowing axial dispersion, which mixes material between neighboring locations. It needs a dispersion parameter as well as the usual flow and elimination inputs. A common representation uses a steady-state convection–dispersion–reaction equation for concentration \(C(z)\) along a liver coordinate \(z\):
Here \(D_{\mathrm{ax}}\) is an axial dispersion coefficient, \(u\) is an effective flow velocity, and \(k\) is an effective first-order elimination rate constant. The equation must be paired with defined inlet and outlet boundary conditions and a mapping from the organ-scale inputs to these parameters. It is not a complete prediction model until those choices are specified. Dispersion can represent behavior between idealized complete mixing and limited axial mixing, but it is not automatically a more accurate model: its extra parameter must be informed by data or defensible prior evidence.
A Short Derivation of the Limiting Predictions
For the well-stirred model, suppose the inlet concentration is \(C_{\mathrm{in}}\) and the mixed liver concentration and outlet concentration are \(C_{\mathrm{out}}\). Under steady state, blood delivers \(Q_h(C_{\mathrm{in}}-C_{\mathrm{out}})\) per unit time, while elimination is \(f_{u,b}CL_{int}C_{\mathrm{out}}\). Equating delivery and elimination gives:
For ideal parallel flow paths, assume concentration declines according to first-order elimination as blood traverses the organ, with the same effective exposure across paths. Integrating the concentration decline over normalized transit gives \(C_{\mathrm{out}}/C_{\mathrm{in}}=\exp(-x)\), so \(E_h=1-\exp(-x)\). These derivations depend on their respective mixing assumptions; they do not establish which assumption applies to a particular liver.
Both formulations also assume that the inputs adequately represent the prediction conditions: flow is specified on the relevant blood basis, unbound fraction and intrinsic clearance are appropriate for those conditions, and the model’s elimination process can be treated as linear over the concentration range of interest. Changes in binding, enzyme activity, transport, or flow can invalidate a prediction even if the mixing assumption is reasonable.
Worked Example: Choosing a Baseline With Limited Mixing Data
A hypothetical PBPK application needs a baseline hepatic clearance prediction. The available inputs are \(Q_h=90\) L/h and \(f_{u,b}CL_{int}=27\) L/h. There are no spatial concentration measurements, transit-time data, or liver-specific evidence about axial mixing. Calculate the two limiting predictions and decide what the evidence supports.
First, \(x=27/90=0.3000\). The well-stirred prediction is:
The parallel-tube prediction is:
The predictions differ by about 2.56 L/h, despite using the same inputs. The available information does not identify a mixing model or estimate a dispersion parameter. A well-stirred model is a defensible parsimonious baseline if its well-mixed assumption is acceptable for the intended application; the parallel-tube result is useful as a structural sensitivity comparison. The calculations alone do not prove that either model describes the liver better.
Match the Formulation to the Task and Evidence
Start by defining what the PBPK application must predict. A model chosen for a routine baseline prediction with sparse inputs may reasonably differ from one used to interpret measured concentration profiles or to investigate a flow perturbation. Then ask what information is available to distinguish the mixing assumptions. In particular, a single clearance measurement may constrain overall extraction without revealing how concentration varies within the organ.
State whether the target is hepatic clearance, extraction, or a change under altered flow, and define the blood, binding, and concentration conditions to which the prediction applies.
Use relevant physiological or experimental information about mixing and transit. Do not treat a clearance value by itself as a direct measurement of intrahepatic mixing.
A dispersion model requires information that can constrain its additional parameter, such as suitable spatial or transit-related measurements, or a justified external basis. Otherwise the extra flexibility may be unidentifiable.
Calculate predictions under plausible formulations using consistent inputs. If the decision or conclusion changes across models, report that structural uncertainty and seek discriminating evidence where practical.
| Formulation | Mixing representation | Useful when | Main caution |
|---|---|---|---|
| Well-stirred | Rapid, organ-wide mixing; one effective concentration | A parsimonious baseline is needed and the assumption is reasonable | Can misrepresent gradients or transit behavior if mixing is not effectively complete |
| Parallel-tube | Concentration falls along parallel paths with little axial back-mixing | Limited axial mixing is supported or is an informative limiting case | Uniform-path assumptions may omit heterogeneity and cross-path mixing |
| Dispersion | Convection, elimination, and axial mixing are represented explicitly | Data or prior evidence can inform the extra mixing parameter | More parameters and boundary-condition choices can create weak identification |
Worked Example: A Prediction Under Limited Axial Mixing
In a hypothetical application, the task is to predict hepatic extraction under conditions for which limited axial back-mixing is considered a plausible physiological representation. The specified inputs are \(Q_h=75\) L/h and \(f_{u,b}CL_{int}=60\) L/h. For comparison, calculate well-stirred and parallel-tube predictions before choosing a working formulation.
The clearance-to-flow ratio is \(x=60/75=0.8000\). Under the parallel-tube assumption:
Under the well-stirred assumption:
The roughly 7.97 L/h difference makes the mixing choice consequential for this prediction. Given the stated evidence favoring limited axial mixing, the parallel-tube formulation is the better-supported choice among these options, provided its parallel-path assumptions are also acceptable. The well-stirred value remains a useful sensitivity comparison; it is not evidence against the selected model by itself.
Worked Example: When Profile Data Support a Dispersion Formulation
A hypothetical study provides measurements at three conditions, summarized by \(x\) and observed \(E_h\). A dispersion model is fitted to profile information using specified boundary conditions. For illustration, suppose its fitted predictions and the two limiting-model predictions are as shown below. The numerical predictions are hypothetical model outputs, not universal values for a given \(x\).
| \(x\) | Observed \(E_h\) | Well-stirred prediction | Parallel-tube prediction | Fitted dispersion prediction |
|---|---|---|---|---|
| 0.300 | 0.245 | 0.2308 | 0.2592 | 0.244 |
| 0.800 | 0.500 | 0.4444 | 0.5507 | 0.501 |
| 1.500 | 0.680 | 0.6000 | 0.7769 | 0.679 |
For a simple descriptive comparison, calculate the sum of squared residuals, \(SSE=\sum_i(E_{h,i,\mathrm{observed}}-E_{h,i,\mathrm{predicted}})^2\). The well-stirred residuals are \(0.0142\), \(0.0556\), and \(0.0800\), giving \(SSE\approx0.00969\). The parallel-tube residuals are \(-0.0142\), \(-0.0507\), and \(-0.0969\), giving \(SSE\approx0.0122\). The dispersion residuals are \(0.001\), \(-0.001\), and \(0.001\), giving \(SSE=0.000003\).
These hypothetical results favor the fitted dispersion model’s agreement with the summarized observations. Selection still depends on more than the smallest SSE: check whether the dispersion parameter is identified, whether residuals show systematic patterns, and whether the fit predicts independent conditions. If the profile information was not actually available, these clearance summaries alone would not justify claiming that the dispersion parameter was measured or that the formulation is physiologically correct.
Diagnostics, Failure Modes, and Limits
A chosen model should be checked against the information it was meant to represent. When measurements across flow conditions or concentration profiles are available, compare predictions with those observations and inspect residuals for systematic departures. Residual patterns can indicate that the assumed mixing structure, inputs, or other model components are inadequate. Agreement at one condition is weak evidence for a model expected to extrapolate to changed flow or other physiological conditions.
A key limitation is identifiability. If several combinations of intrinsic clearance and mixing parameters produce nearly the same observed hepatic clearance, the data may not support a precise dispersion estimate. More parameters can improve fit without providing a uniquely supported biological explanation. A dispersion formulation also depends on choices such as boundary conditions and the representation of flow and elimination; these should be stated rather than treated as incidental implementation details.
The three formulations are not a complete catalogue of hepatic physiology. They may not represent heterogeneous flow paths, changing enzyme activity, transporter effects, or concentration-dependent elimination adequately. Their predictions also rely on compatible definitions of flow, binding, and intrinsic clearance. Model selection cannot repair incorrect input scaling or an unmodeled biological process.
Common Mistake: Treating Complexity as Evidence
- Wrong reasoning: “The dispersion model is more detailed, so it must give the best prediction.”
- Correction: Detail is useful only when its assumptions are suitable and its additional parameter is supported by information. With sparse data, a simpler model may be more defensible, and predictions from alternative models can show the size of structural uncertainty.
- Wrong reasoning: “One clearance measurement matches the well-stirred prediction, so the liver is well mixed.”
- Correction: A single overall clearance can be consistent with different internal mixing patterns. Model agreement at one condition does not by itself identify the mixing mechanism.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- For \(Q_h=100\) L/h and \(f_{u,b}CL_{int}=20\) L/h, calculate \(x\), then the well-stirred and parallel-tube extraction ratios.
- A clearance value is available at one flow condition, but no spatial or transit-related data are available. Does that observation alone identify a dispersion parameter? Why?
- State the mixing assumption that distinguishes the well-stirred formulation from the parallel-tube formulation.
- Why might a lower residual sum of squares for a dispersion fit not be enough to justify using it for an extrapolation?
Solutions. (1) \(x=20/100=0.2000\). Well-stirred \(E_h=0.2/1.2=0.1667\); parallel-tube \(E_h=1-\exp(-0.2)=0.1813\). (2) No. One overall clearance value generally does not reveal the internal concentration profile or uniquely determine an additional mixing parameter. (3) The well-stirred model assumes rapid organ-wide mixing and one effective concentration; the parallel-tube model assumes concentration declines along paths with little axial back-mixing. (4) A more flexible fit may reflect extra parameters rather than reliable structure, and may not predict independent conditions. Check parameter identifiability, residual patterns, and performance beyond the fitted data.