Why Ontogeny Uncertainty Matters
A pediatric PBPK prediction depends not only on the selected enzyme or transporter maturation curve, but also on how well that curve is known. Uncertainty in the age at which activity rises, the rate of rise, or the mature level can change predicted clearance and therefore exposure. Because a PBPK model combines these processes nonlinearly, uncertainty in the input curve does not generally translate into a symmetric uncertainty range for exposure.
The aim is to propagate uncertainty in the maturation functions through the model, rather than to attach an arbitrary margin to its final prediction. The result can show both how uncertain exposure is at a given age and where better developmental evidence could have the greatest value. Earlier material on enzyme, transporter, and co-maturation profiles supplies biological context; here the focus is how uncertainty in those profiles moves through simulations.
These intervals are conditional on the model structure, dose, age definition, and the distributions assigned to uncertain inputs. A 95% interval does not mean there is a 95% probability that a fixed true exposure lies inside it. In a probabilistic simulation, it means that 95% of the simulated exposure distribution is between the stated quantiles, given the specified model and input distributions.
Set Up the Uncertainty Before Simulating
Represent an enzyme or transporter maturation function at age \(a\) as \(M(a;\theta)\), where \(\theta\) is a vector of uncertain parameters. For example, a saturating curve might be written as
Here \(T_M\) is the age at which this particular curve reaches one half of its mature level, and \(h\) controls its shape. The equation is an example of a possible curve, not a claim that every developmental profile follows this form. The uncertainty analysis requires defensible distributions for the parameters, parameter bounds consistent with the biology, and a joint distribution when parameters or pathways are dependent.
For several enzymes and transporters, collect the uncertain parameters into a joint vector \(\Theta\). A PBPK simulation at age \(a\) then maps a draw of these parameters to an exposure outcome:
where \(g\) denotes the PBPK calculation and \(\Phi\) represents other inputs, such as physiological characteristics, dose, and drug-specific properties. If the question is specifically about ontogeny uncertainty, keep other inputs fixed or state clearly which of them are also sampled. If individual variability is to be included, it must be represented explicitly and not confused with uncertainty about the average maturation curve.
A key assumption is that the assigned input distributions capture the uncertainty relevant to the question. Sampling independent marginal distributions is not automatically justified: shared developmental biology can make enzyme and transporter maturation parameters dependent. The relationship between parameters should be represented in the joint sampling scheme, or the independence assumption should be defended. Draws should also respect meaningful bounds, such as a maturation fraction between zero and one where that definition applies.
Worked Example: Transforming a Maturation Range Into an Exposure Interval
Consider a deliberately simplified, hypothetical exposure calculation at one pediatric age. Let the maturation fraction \(M\) for a clearance pathway be uniformly distributed from 0.10 to 0.30. Let the model output be
The constants summarize fixed dose and other clearance contributions for this example; the only uncertainty being propagated is \(M\). Because the denominator increases with \(M\), AUC decreases with \(M\). Thus the lower AUC endpoint corresponds to the upper maturation quantile, not the lower one.
The median maturation fraction is 0.20, so the median AUC is \(100/[1+2(0.20)]=100/1.40=71.43\ \mathrm{mg\cdot h/L}\). For a uniform distribution on 0.10 to 0.30, the 2.5th and 97.5th percentiles of \(M\) are 0.105 and 0.295. Transforming these in reverse order gives \(100/[1+2(0.295)]=100/1.59=62.89\ \mathrm{mg\cdot h/L}\) and \(100/[1+2(0.105)]=100/1.21=82.64\ \mathrm{mg\cdot h/L}\). The central 95% interval is therefore approximately 62.89 to 82.64 mg·h/L.
For completeness, the mean AUC is not the AUC evaluated at the mean \(M\). Direct integration gives \(\mathrm{E}(\mathrm{AUC})=(1/0.20)\int_{0.10}^{0.30}100/(1+2m)\,dm =250\ln(4/3)=71.92\ \mathrm{mg\cdot h/L}\). The median and mean differ because the transformation is nonlinear. The interval is valid for this example because the assumed distribution and monotone model map have been stated explicitly.
Monte Carlo Propagation Through a PBPK Model
For a full PBPK model, the map from maturation parameters to exposure is usually too complicated to transform by hand. Monte Carlo propagation provides a direct approach: draw a plausible joint set of uncertain inputs, run the model, and summarize the resulting exposure values. The same input draw should be used consistently wherever that parameter enters the simulation.
Choose the age, dose, exposure metric, and which uncertainties are included. For example, define the target as AUC after a stated dose at a specified postnatal age, with ontogeny parameter uncertainty only.
Assign distributions and bounds to uncertain maturation parameters, including dependence where supported. Document fixed inputs and assumptions about any unrepresented uncertainties.
For each draw \(i\), calculate the relevant age-specific maturation functions and run the PBPK model to obtain \(Y_i(a)\). Repeat for enough draws to estimate the summaries of interest stably.
Report the median and chosen quantiles, in context and with units. Check quantile stability across larger runs or repeated batches, and inspect whether unusual outputs trace to plausible inputs or to model failures.
The empirical 2.5th and 97.5th percentiles of the simulated values form a central 95% simulation interval. This interval need not be symmetric about the median. Its meaning depends on what was sampled: sampling only uncertain ontogeny parameters yields an uncertainty interval conditional on the other inputs, not a complete interval for individual children.
Worked Example: Covariance-Aware Local Approximation
A hypothetical local model for AUC is \(Y=100/(1+2M_E+M_T)\ \mathrm{mg\cdot h/L}\), where \(M_E\) and \(M_T\) are enzyme and transporter maturation fractions. Suppose their means are 0.20 and 0.30, their standard deviations are 0.05 and 0.04, and their correlation is 0.75. This approximation is intended for relatively small variation around the means; it is not a replacement for full simulation when curvature or bounds matter.
At the mean inputs, the denominator is \(1+2(0.20)+0.30=1.70\), so \(Y=100/1.70=58.82\ \mathrm{mg\cdot h/L}\). The first-order delta method approximates output variance by the variance of the linear expansion around these means. The required derivatives are \(\partial Y/\partial M_E=-200/(1+2M_E+M_T)^2=-69.20\) and \(\partial Y/\partial M_T=-100/(1+2M_E+M_T)^2=-34.60\), evaluated at the means.
The input covariance is \(0.75(0.05)(0.04)=0.0015\). Therefore,
The approximate standard deviation is \(\sqrt{21.07}=4.590\ \mathrm{mg\cdot h/L}\). If the local output distribution is treated as approximately normal, a rough central 95% interval is \(58.82\pm1.96(4.590)\), or 49.83 to 67.82 mg·h/L. The positive covariance increases the approximate variance because both derivatives are negative, so joint increases in maturation tend to lower AUC together. The interval is only a local approximation: a full simulation is preferable if the distributions are bounded, skewed, or sufficiently broad that the nonlinear model changes the output shape.
Age-Specific Intervals and Evidence Priorities
An uncertainty interval should be calculated separately at each age of interest. The same uncertain maturation curve can have very different consequences at different ages: a parameter shift may strongly change activity during a steep developmental transition but have little effect when the curve is near its mature plateau. When comparing ages, use a coherent joint draw of the curve parameters across ages. Otherwise, apparent differences may reflect independently resampled curves rather than developmental uncertainty in a single plausible trajectory.
Worked Example: Locating the Age With the Widest Exposure Forecast
Suppose a hypothetical PBPK analysis draws 10,000 joint enzyme and transporter parameter sets and simulates AUC at two ages after the same dose. The resulting quantiles are:
| Age | 2.5th percentile | Median | 97.5th percentile | Interval width |
|---|---|---|---|---|
| 1 month | 42 mg·h/L | 68 mg·h/L | 112 mg·h/L | 70 mg·h/L |
| 5 years | 51 mg·h/L | 76 mg·h/L | 108 mg·h/L | 57 mg·h/L |
At 1 month, the central 95% interval is 42 to 112 mg·h/L, with width \(112-42=70\ \mathrm{mg\cdot h/L}\). At 5 years, it is 51 to 108 mg·h/L, with width \(108-51=57\ \mathrm{mg\cdot h/L}\). Under the stated input distributions, the younger-age forecast is more dispersed, although its median is lower. The intervals overlap, so these summaries alone do not establish that exposure at 1 month is always lower than exposure at 5 years.
To investigate which evidence most limits the 1-month forecast, suppose diagnostic reruns yield a width of 28 mg·h/L when enzyme uncertainty is fixed at its conditional median while transporter uncertainty remains, and 61 mg·h/L when transporter uncertainty is fixed while enzyme uncertainty remains. In this hypothetical analysis, removing enzyme uncertainty reduces interval width more. That makes enzyme maturation evidence a priority for follow-up. These reruns are a sensitivity diagnostic, not an additive variance decomposition; dependence between the inputs can affect the comparison, so the conditioning and sampling scheme must be reported.
Diagnostics, Alternatives, and Limits
Several checks help determine whether the propagated interval is credible. First, inspect sampled parameter values and age-specific maturation curves, not just exposure outputs. Unexpectedly extreme curves can signal poorly chosen distributions, mistaken parameter units, or an unmodelled dependence. Second, check simulation failures and the response of key outputs to input changes. If greater maturation of a clearance pathway increases AUC in a model where it should increase clearance, verify the implementation and pathway interpretation. Third, compare quantiles across repeated batches or larger simulation runs; unstable tails may indicate too few effective draws, especially when inputs are correlated or outputs are highly skewed.
Monte Carlo simulation is flexible for nonlinear models and non-normal outputs, but it does not repair weak evidence or a misspecified model. A local delta-method approximation is quicker and can reveal how parameter variances and covariances affect output uncertainty, but it can fail for broad uncertainty, strong curvature, or boundary-constrained inputs. Scenario analysis is useful for transparent best-case and worst-case assumptions, but its endpoints are not probability intervals unless scenarios and weights have a defensible probabilistic interpretation. These approaches answer related but different questions.
Uncertainty in the maturation curve is not the same as uncertainty about the model structure. For example, a simulation that samples \(T_M\) and \(h\) under one curve family does not account for uncertainty about whether that family is appropriate. Nor does it automatically include uncertainty in organ size, binding, dose history, or other physiological inputs. State these omissions, and avoid presenting a narrow interval as comprehensive when important sources were held fixed. If parameters were estimated from sparse age-specific evidence, the resulting interval is conditional on the evidence model and its assumptions.
Common Pitfall
A second frequent error is to call an ontogeny-only interval a pediatric population prediction interval. The former describes how parameter uncertainty moves a model prediction; the latter also requires a defined representation of variability between children. Label the uncertainty source, age, dose, exposure metric, and interval construction so readers can interpret the forecast correctly.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If AUC decreases monotonically as maturation increases, which maturation percentile maps to the lower endpoint of the central 95% AUC interval?
- In a simulation that samples only maturation parameters, what does the resulting 95% interval describe, and what does it not automatically describe?
- Why can independent sampling of enzyme and transporter parameters be misleading?
- In the local example, why does positive covariance increase approximate AUC variance when both AUC derivatives are negative?
- A 1-month interval has width 70 mg·h/L and a 5-year interval has width 57 mg·h/L. What can be concluded, and what cannot be concluded from widths alone?
Solutions. (1) The 97.5th maturation percentile maps to the lower AUC endpoint because the transformation is decreasing. (2) It describes uncertainty in the model output conditional on the sampled maturation distributions and fixed inputs; it does not automatically describe between-child variability or a complete population prediction interval. (3) Dependence can make some parameter combinations more or less plausible, changing the distribution of simulated exposure. (4) The covariance contribution is \(2(\partial Y/\partial M_E)(\partial Y/\partial M_T)\mathrm{Cov}(M_E,M_T)\); both derivatives and the covariance are negative, negative, and positive respectively, so the term is positive. (5) The younger-age forecast is more dispersed under the specified simulation assumptions. The widths alone do not establish which age has higher exposure for every simulated child, nor do they establish a causal age effect.