Why the Source of a Prediction Error Matters
When a PBPK model does not match concentration data, changing a parameter is not always the right response. A partitioning value or exchange rate may be poorly chosen, but the model may instead connect or combine tissues in a way that cannot reproduce the observed time course. These explanations can look similar at selected times, yet they imply different model changes.
The distinction is practical. If the compartment arrangement is adequate, revising a parameter may improve predictions across the relevant time course. If the arrangement is inadequate, parameter adjustment may improve one part of the profile while leaving systematic discrepancies elsewhere. The aim here is not to declare a structure wrong from one residual, but to compare the patterns predicted by specific alternatives.
The earlier material on reading structural signals in concentration profiles provides the context for recognizing profile patterns. The central task here is to ask whether a plausible change in a parameter within the existing structure can explain the pattern, or whether the structure itself restricts the shapes the model can produce.
What Different Errors Tend to Do
A parameter change acts within the prediction family allowed by a chosen structure. For example, in a simple exchange model, changing the exchange-rate parameter changes how quickly the tissue approaches its equilibrium concentration. If the model has only one exchange timescale, changing that rate can make the approach faster or slower, but it cannot create two distinct timescales.
A structural change can alter the family of possible profiles. Separating a fast-exchanging tissue from a slow-exchanging tissue may produce an initial rapid change followed by a slower approach. A single effective compartment may be unable to reproduce both phases with one fixed exchange rate. This is a diagnostic clue, not a guarantee: sparse sampling, measurement noise, or a limited observation window can hide the pattern.
Worked Example: A Wrong Exchange Rate in a Fixed Structure
Consider a hypothetical one-compartment exchange model with tissue concentration \(C(t)=C_{\mathrm{eq}}(1-e^{-kt})\), where \(C_{\mathrm{eq}}\) is the equilibrium concentration and \(k\) is an exchange-rate constant in \(\mathrm{h}^{-1}\). Assume the initial tissue concentration is zero, \(C_{\mathrm{eq}}=12.0\ \mathrm{mg/L}\), and the data-generating rate is \(k=0.300\ \mathrm{h}^{-1}\). Compare it with the same model using a slower value, \(k=0.150\ \mathrm{h}^{-1}\).
At \(t=2.00\ \mathrm{h}\), the data-generating concentration is \(12.0(1-e^{-0.300(2.00)})=12.0(1-e^{-0.600})\approx 5.414\ \mathrm{mg/L}\).
The slower-rate model predicts \(12.0(1-e^{-0.150(2.00)})=12.0(1-e^{-0.300})\approx 3.110\ \mathrm{mg/L}\). The prediction is therefore approximately \(5.414-3.110=2.304\ \mathrm{mg/L}\) too low at this time.
Here the two predictions differ because the rate parameter differs; both use the same single-compartment structure and the same equilibrium concentration. If the slower model were too low over the rising part of a profile, increasing \(k\) would be a reasonable parameter adjustment to investigate. It would not, by itself, establish that the single-compartment structure is adequate.
Why One Rate May Not Represent Two Kinetic Processes
A simple derivation shows why distinct exchange rates can produce a profile that one rate cannot match throughout. Suppose the observed concentration is a weighted combination of two tissue signals. Both begin at zero and approach the same normalized equilibrium value \(C_{\mathrm{eq}}\), but their exchange rates are \(k_1\) and \(k_2\). Let \(f\) and \(1-f\) be their respective contributions, with \(0<f<1\). The combined concentration is
The fraction of the equilibrium concentration not yet reached is \(R(t)=1-C(t)/C_{\mathrm{eq}}=fe^{-k_1t}+(1-f)e^{-k_2t}\). A single-exponential model would require \(R(t)=e^{-kt}\) for one fixed \(k\). Taking the logarithm defines the rate that a single exponential would need at each time:
If \(k_1\ne k_2\), the weighted sum of exponentials is generally not itself a single exponential, so \(k_{\mathrm{app}}(t)\) need not remain constant. In fact, the initial slope gives a weighted average rate: differentiating \(C(t)\) at \(t=0\) gives \(C'(0)=C_{\mathrm{eq}}[fk_1+(1-f)k_2]\). At later times, the slower component contributes a larger share of the remaining unfilled concentration, so the apparent rate tends toward the slower rate. This is a mathematical reason that a single rate may fit one time region and miss another.
Worked Example: A Structural Alternative With Fast and Slow Exchange
Suppose hypothetical data are generated by two tissue contributions with \(C_{\mathrm{eq}}=10.0\ \mathrm{mg/L}\), \(f=0.600\), \(k_1=0.800\ \mathrm{h}^{-1}\), and \(k_2=0.100\ \mathrm{h}^{-1}\). The model therefore predicts \(C(t)=10.0[1-0.600e^{-0.800t}-0.400e^{-0.100t}]\).
At \(t=1.00\ \mathrm{h}\), the remaining fraction is \(0.600e^{-0.800}+0.400e^{-0.100} =0.600(0.449329)+0.400(0.904837) \approx 0.631532\). Thus \(C(1.00)\approx10.0(1-0.631532)=3.685\ \mathrm{mg/L}\), and the rate required for a single exponential to match this time is \(k_{\mathrm{app}}=-\ln(0.631532)/1.00\approx0.4596\ \mathrm{h}^{-1}\).
At \(t=3.00\ \mathrm{h}\), the remaining fraction is \(0.600e^{-2.400}+0.400e^{-0.300} =0.600(0.090718)+0.400(0.740818) \approx0.350758\). Thus \(C(3.00)\approx10.0(1-0.350758)=6.492\ \mathrm{mg/L}\), and the single-exponential rate required at this time is \(k_{\mathrm{app}}=-\ln(0.350758)/3.00\approx0.3492\ \mathrm{h}^{-1}\).
The two times require different apparent rates, approximately \(0.4596\) and \(0.3492\ \mathrm{h}^{-1}\). Consequently, a single fixed exchange rate cannot match both concentrations exactly. This discrepancy is consistent with the two-rate structure used to generate the data; it does not establish that real tissue kinetics must have exactly two compartments or these particular rates.
When Concentration Data Cannot Separate the Explanations
Different structures can make identical predictions for a particular observable. If two proposed models give the same concentration at every measured time under the conditions studied, those measurements alone cannot distinguish them. This is not evidence that the structures are physiologically identical; it means only that the chosen observations do not reveal their differences.
For example, consider a normalized measured uptake signal. One model represents a single tissue pool with \(y(t)=1-e^{-0.400t}\). Another represents two pools contributing equally, each with the same exchange rate \(0.400\ \mathrm{h}^{-1}\). Its measured signal is \(y(t)=0.5(1-e^{-0.400t})+0.5(1-e^{-0.400t})=1-e^{-0.400t}\). The predictions agree at all times, even though one representation has one pool and the other has two. A concentration time course of this signal cannot choose between them.
More generally, ambiguity can arise when the measured concentration is an aggregate, when observations cover only a narrow time range, or when the parameters of competing structures compensate for one another. Measurements from additional tissues, a different dosing condition, or a wider time course may provide new information, but whether they do so depends on the model and observation process. A new measurement is useful only if the candidate structures predict meaningfully different values for it.
Worked Example: Identical Predictions From Different Arrangements
Suppose two hypothetical models predict a normalized tissue signal at \(t=2.00\ \mathrm{h}\). Model A has one exchange pool with rate \(0.400\ \mathrm{h}^{-1}\). Model B has two equally weighted pools, each with that same rate. Under the stated assumptions, Model A predicts \(y_A=1-e^{-0.400(2.00)}=1-e^{-0.800}\approx0.550671\).
Model B predicts \(y_B=0.5(1-e^{-0.800})+0.5(1-e^{-0.800})=1-e^{-0.800}\approx0.550671\). The predictions are identical, and the equality holds at every time \(t\), not just at two hours, because the weighted terms are identical. Even error-free measurements of this signal cannot distinguish these two representations. A measurement that resolves the pools separately, or a condition under which their rates differ, could potentially separate them.
Practical Diagnosis, Comparisons, and Limitations
A disciplined comparison starts with the current structure and asks what it can generate as its parameters vary. For a one-rate approach-to-equilibrium model, different rates change the speed of approach, while the equilibrium value changes the limiting concentration. If the observed profile has a plausible common equilibrium but its early and late portions imply incompatible rates, changing only the rate is unlikely to resolve the entire discrepancy. A structure with multiple kinetic components is one possible explanation to test.
A useful diagnostic is to examine prediction errors across time, rather than focusing on one concentration. Residuals that are persistently positive in one time region and negative in another can indicate a shape mismatch, though sampling variability and measurement error can produce irregular patterns. Check whether the mismatch remains when plausible parameter values are considered, whether equilibrium or dose-related assumptions are also at issue, and whether the observation data actually distinguish the proposed structures.
Record whether predictions are high or low and identify the time regions involved. Do not infer a structural problem from one isolated difference.
Ask whether a plausible change in partitioning, exchange, or another relevant parameter could improve the profile without creating a new systematic mismatch elsewhere.
Consider a different compartment arrangement only when it represents a plausible process and can produce a prediction pattern unavailable to the current structure.
Determine whether the structures predict different measured concentrations under the available conditions. If predictions coincide, these data cannot resolve the structural choice.
The comparison is not a formal proof of model truth. A parameter adjustment is usually simpler to interpret and preserves the existing arrangement, but it cannot overcome a genuine restriction in the model’s possible profile shapes. A structural alternative may represent distinct exchange processes, but extra compartments introduce additional parameters and assumptions. With limited or noisy data, those additions may not be estimable and can make predictions less stable.
Concentration profiles also do not uniquely identify physiological mechanisms in every setting. Similar curves can arise from different combinations of structure and parameter values, and an apparent shape discrepancy may reflect an incorrect input, measurement timing, or another assumption rather than compartment arrangement. Treat structural diagnosis as a comparison of plausible explanations, and state what the observations do and do not support.
Common Pitfall
- Wrong reasoning: “The model is too low at two hours, so the structure must be wrong.” A single mismatch does not distinguish a wrong exchange rate from a wrong arrangement, and may also reflect other assumptions or measurement error.
- Correction: Compare the full time pattern. Check whether a plausible parameter change within the current structure can explain the early and late observations together. If no single value can do so, test a physiologically plausible structural alternative, then assess whether the available observations can distinguish the models.
Check Your Understanding
Answer each question using the distinction between parameter values and model arrangement. Brief solutions follow the questions.
- A single-compartment model predicts concentrations that are consistently too low during the rising phase, but approaches the observed equilibrium concentration. What parameter change is a reasonable first explanation, and why does it not prove the structure is adequate?
- A one-rate model requires an apparent rate of \(0.50\ \mathrm{h}^{-1}\) to match an early observation and \(0.20\ \mathrm{h}^{-1}\) to match a later observation. What does this suggest, and what does it not prove?
- Two structural models give exactly the same predicted measured concentration at every observed time. Can those observations select one structure? Explain.
- Why might adding a second tissue compartment fail to improve reliable inference even if it reduces the apparent mismatch in a fitted profile?
Solutions. (1) Increasing the exchange rate is a reasonable parameter explanation because it speeds approach while leaving the limiting concentration unchanged in this simple model. It does not prove adequacy because another time region may remain mismatched, and the data may not reveal the underlying arrangement. (2) The incompatible rates suggest that one fixed rate may not represent the whole profile and motivate testing a structural alternative. They do not prove that the system has two compartments; noise, sparse sampling, or other incorrect assumptions can produce a similar pattern. (3) No. If the models predict the same measured values under the observed conditions, those observations cannot distinguish them. (4) The extra compartment adds assumptions and parameters that may be weakly identified by the available data, so estimates and predictions can become unstable despite a closer fit.