Why residual patterns matter
An age-stratified PBPK model can produce plausible average exposure predictions while systematically missing the time course in a particular tissue or age group. A single summary, such as an overall exposure ratio, can conceal whether the model predicts too much concentration early and too little later, or whether it misses concentrations in one tissue but not others. Those patterns help assess whether age-related organ volume or blood flow is a plausible source of error.
A residual is a difference between an observation and a model prediction, defined on a stated scale and at a matched time and sampling matrix. Residuals are diagnostic clues, not direct measurements of physiology. The same pattern can arise from more than one model input, and observed data may also reflect measurement error, sampling differences, or an incorrect structural assumption.
For example, \(r=0.20\) corresponds to an observed-to-predicted ratio of \(e^{0.20}=1.221\), so the observed concentration is about 22.1% above prediction. A residual near zero indicates agreement at that observation, not that all model components are correct.
How volume and flow shape tissue time courses
A useful starting point is a perfusion-limited tissue with tissue volume \(V_t\), blood flow \(Q_t\), arterial concentration \(C_a\), tissue concentration \(C_t\), and tissue-to-blood partition coefficient \(K_{p,t}\). Assume well-mixed blood and tissue, instantaneous partition equilibrium between venous blood and tissue, no tissue metabolism, and constant \(C_a\), \(Q_t\), \(V_t\), and \(K_{p,t}\) over the interval considered. Under these simplifying assumptions, the tissue amount is \(A_t=V_tC_t\), and the venous concentration is \(C_{v,t}=C_t/K_{p,t}\).
Flow conservation gives the rate of change in tissue amount as inflow minus outflow. Substituting \(A_t=V_tC_t\) and \(C_{v,t}=C_t/K_{p,t}\) yields the following relationship:
Rearranging gives a first-order approach to equilibrium. If \(C_t(0)=0\) and \(C_a\) is constant, the solution is \(C_t(t)=K_{p,t}C_a\left(1-e^{-t/\tau_t}\right)\), where \(\tau_t=V_tK_{p,t}/Q_t\). This time constant has units of time: volume multiplied by a dimensionless partition coefficient and divided by volume per time.
Thus, increasing \(V_t\) or decreasing \(Q_t\) increases \(\tau_t\) in this model and slows the approach to equilibrium. The limiting concentration \(K_{p,t}C_a\) does not depend on flow or volume under these assumptions. In a full PBPK model, however, \(C_a\) itself may change as volumes, flows, and other organs interact. Organ volume also changes tissue capacity: at a given tissue concentration, the amount \(A_t=V_tC_t\) is proportional to volume. These equations explain why a time-dependent, tissue-specific residual pattern can be more informative than a single exposure summary, while also showing why it cannot by itself identify a unique physiological cause.
Worked Example: Reading a Time-Crossing Tissue Residual
Suppose a hypothetical older-adult dataset reports liver concentrations at three times after dosing. The PBPK predictions are 10.0 mg/L at 1 hour, 6.00 mg/L at 6 hours, and 2.00 mg/L at 18 hours. The corresponding observations are 8.00, 6.00, and 3.00 mg/L. Calculate the log residuals and interpret their pattern.
At 1 hour, \(r=\log(8.00/10.0)=\log(0.800)=-0.223\). The model overpredicts by 20.0% relative to its prediction.
At 6 hours, \(r=\log(6.00/6.00)=0\), so prediction and observation agree at that time.
At 18 hours, \(r=\log(3.00/2.00)=\log(1.50)=0.405\). The observation is 50.0% above prediction. The residual changes from negative early to positive late: the model predicts a higher early concentration and a lower late concentration than observed.
If arterial or plasma predictions fit well over the same interval, a local tissue-distribution mismatch becomes more plausible than a broad systemic input error. In the simplified equation above, the crossing pattern is compatible with the model approaching tissue equilibrium too quickly; a lower tissue flow or larger tissue volume could lengthen the predicted time constant. It does not prove either explanation: partitioning, tissue exchange assumptions, and errors in the blood concentration driving the tissue can also matter.
Compare patterns across tissues and age groups
Interpret residuals in matched strata: the same dose basis, sampling times, matrix, and exposure definition. Compare the age-specific pattern with the model's age-specific predictions rather than assuming that every older individual should have the same physiology. Earlier material on age-related organ atrophy and blood flow provides physiological context; here the question is whether the model's predicted exposure errors are compatible with those inputs.
A broadly similar residual across plasma and several tissues may point toward a systemic driver, such as an input or clearance discrepancy, rather than a single organ's volume or perfusion. By contrast, a residual concentrated in a particular tissue, especially one that changes systematically over time while plasma is reasonably predicted, is more consistent with a local distribution problem. Neither pattern is unique to one mechanism.
| Residual pattern | What it may be compatible with | What to check before interpreting |
|---|---|---|
| Early and late tissue residuals have opposite signs; plasma fits reasonably | Incorrect tissue equilibration rate, potentially related to flow or volume | Sampling times, tissue-specific partitioning, and tissue concentration measurement |
| Similar signed residuals in plasma and multiple tissues | A systemic prediction error affecting tissue delivery | Dose, absorption or input, clearance, and consistency of the prediction target |
| Persistent residual in one tissue with little time variation | Incorrect partitioning, capacity, or a tissue-specific model assumption | Whether the data are near equilibrium and whether the tissue matrix is comparable |
| Residuals vary substantially among people of the same age | Unrepresented physiological variation or noisy observations | Cohort composition, covariates, and the amount of data supporting each stratum |
A residual trend against time is not automatically evidence of age-related physiology. Check whether the pattern also appears in younger participants, whether it is driven by a few observations, and whether prediction uncertainty and measurement variability are large enough to explain it. When repeated observations are available, inspect the individual time courses as well as an age-group summary; averaging can hide opposite patterns among individuals.
Worked Example: Separating a Tissue-Specific Pattern From a Broad Offset
In a hypothetical late-time comparison for older adults, suppose predicted and observed concentrations are 5.00 and 5.20 mg/L in kidney, 2.00 and 1.40 mg/L in muscle, and 3.00 and 4.50 mg/L in adipose tissue. Calculate each log residual.
For kidney, \(r=\log(5.20/5.00)=\log(1.04)=0.0392\), a 4.00% relative underprediction. For muscle, \(r=\log(1.40/2.00)=\log(0.700)=-0.357\), a 30.0% relative overprediction. For adipose tissue, \(r=\log(4.50/3.00)=\log(1.50)=0.405\), a 50.0% relative underprediction.
These residuals do not show a common offset across all three tissues: kidney is close to prediction, muscle is overpredicted, and adipose is underpredicted. That pattern is not well summarized as a single age-related change in systemic exposure. It motivates checking tissue-specific assumptions and measurements, including partitioning and perfusion, but cannot select among them. Because these are late-time values, they also do not establish a difference in equilibration rate without earlier observations.
Use counterfactual simulations to assess plausibility
A practical diagnostic is to change one physiological input at a time, within independently defensible bounds, and compare the resulting concentration-time predictions with the residual pattern. Such a counterfactual is a sensitivity check: it asks whether the proposed change moves predictions in the needed direction and at the needed times. It does not establish that the changed input is the true cause.
Worked Example: Why a Single Time Point Cannot Separate Flow From Volume
Consider the simplified tissue model above with constant arterial concentration \(C_a=10.0\) mg/L, \(K_{p,t}=2.00\), and initial tissue concentration zero. A baseline model has \(Q_t=0.600\) L/min and \(V_t=3.00\) L. Its time constant is \(3.00(2.00)/0.600=10.0\) minutes. At 10.0 minutes, it predicts \(C_t=20.0(1-e^{-1})=12.64\) mg/L.
Suppose the hypothetical observed concentration at 10.0 minutes is 7.87 mg/L. The baseline residual is \(r=\log(7.87/12.64)=-0.474\), indicating overprediction. Now reduce flow to 0.300 L/min while leaving volume at 3.00 L. The time constant becomes \(3.00(2.00)/0.300=20.0\) minutes, and the prediction is \(20.0(1-e^{-10/20})=7.869\) mg/L.
Alternatively, keep flow at 0.600 L/min and increase volume to 6.00 L. The time constant is also \(6.00(2.00)/0.600=20.0\) minutes, producing the same 7.869 mg/L prediction at 10.0 minutes. Both changes remove the residual at this time point, but they imply different physiology. In this simplified model, they produce the same entire concentration-time curve because both give the same ratio \(V_tK_{p,t}/Q_t\). A concentration observation alone therefore cannot distinguish these alternatives; independent age-specific flow or volume information, and evaluation of the wider model, are needed.
When running counterfactuals in a complete PBPK model, check that the change is physiologically coherent across organs and that cardiac output and flow allocation remain consistent. Also inspect whether the modification improves the target tissue without worsening plasma or other tissues. If a proposed volume or flow change only helps after implausible values are used, it is a weak explanation even if the residual decreases.
Common pitfall and limits
Wrong reasoning: “Older participants have negative early tissue residuals and positive late residuals, so reduced blood flow is the cause.” This overstates what the pattern establishes. In the simplified model, both lower flow and larger volume can slow equilibration, and other assumptions can produce similar residuals.
Correction: State the observation first: the model's tissue time course appears too fast relative to the data, conditional on the model and measurements. Then test whether age-specific flow and volume values are independently plausible, use counterfactual simulations, and check competing explanations such as partitioning or systemic concentration errors. The residual supports a diagnostic hypothesis; it does not identify causation.
Residual analysis also has limits when data are sparse, tissue samples are not comparable across ages, or a model prediction is evaluated at times far from the actual sampling schedule. A population-average residual can mask person-level heterogeneity. A close fit in one tissue does not validate the full physiology, and a poor fit does not by itself establish that organ volume or blood flow is wrong. The interpretation should remain tied to the data, model assumptions, and uncertainty.
Check Your Understanding
Use the sign convention \(r=\log(C_{\mathrm{obs}}/C_{\mathrm{pred}})\). Solutions follow each question.
- A prediction is 4.00 mg/L and the observation is 5.00 mg/L. Find the residual and state whether the model underpredicts or overpredicts.
Solution: \(r=\log(5.00/4.00)=\log(1.25)=0.223\). The positive residual means the model underpredicts; the observation is 25.0% above prediction. - Under the simplified perfusion-limited model, what happens to \(\tau_t\) if \(V_t\) doubles while \(Q_t\) and \(K_{p,t}\) remain fixed?
Solution: Since \(\tau_t=V_tK_{p,t}/Q_t\), doubling \(V_t\) doubles \(\tau_t\). The tissue approaches its limiting concentration more slowly. - Plasma residuals are near zero, while an early tissue residual is negative and a late tissue residual is positive. What is one plausible interpretation, and why is it not conclusive?
Solution: The model may predict tissue equilibration too quickly. Lower flow or larger tissue volume could lengthen the time constant, but partitioning, tissue exchange assumptions, or other errors could also explain the pattern. - Why can a single tissue concentration time point be insufficient to distinguish lower blood flow from larger organ volume?
Solution: In the simplified model, the time course depends on \(\tau_t=V_tK_{p,t}/Q_t\). Different volume and flow values can yield the same time constant and thus the same predicted concentration, so independent physiology data or additional constraints are needed.