Mapping Input Uncertainty to PBPK Prediction Distributions

From Inputs to a Prediction Distribution

A PBPK model can produce a concentration-time profile for one specified set of inputs. But many practical questions concern a range of plausible predictions: how much might oral-dose AUC vary under input uncertainty, or what concentrations might occur in a defined population at a specified time? Answering these questions requires more than running the model once. It requires stating which inputs are uncertain, how they are represented, what output is being predicted, and whose predictions are relevant.

The central operation is a mapping: an input vector is passed through a PBPK model, and the model returns an output. If the inputs have a specified distribution, applying that mapping induces a distribution of outputs. This distribution is meaningful only relative to the input specification, the model, the regimen, the endpoint, and the target population.

Definition: Let \(\mathbf{X}\) denote the vector of PBPK inputs treated as uncertain, and let \(G_T\) be their specified joint distribution for a stated target \(T\). For a fixed regimen and model structure, let \(f\) map an input vector to a scalar endpoint. The prediction is the random variable \(Y=f(\mathbf{X})\). Its distribution is the prediction distribution induced by applying \(f\) to \(G_T\). The target \(T\) states the population or parameter context to which the input distribution and prediction refer.

Examples of inputs include physiological quantities, drug-specific parameters, and covariates used by the model. The output might be AUC, concentration at a specified time, or another defined endpoint. A distribution for inputs is not merely a list of ranges: probabilities or weights must be specified if the resulting output distribution is to have a probabilistic interpretation. How to select input distributions is a separate modeling decision; here the focus is how a stated input distribution maps to predictions.

What the Prediction Distribution Means

Suppose \(G_T\) assigns probability to possible input vectors. The model maps each possible vector to an output value. For any threshold \(y\), the output cumulative distribution is the probability that the input falls in the set of values whose model output is no greater than \(y\):

P(Y \le y) = P\bigl(f(\mathbf{X}) \le y\bigr)

This relation is the basic justification for propagating input uncertainty through a PBPK model. It does not require \(Y\) to have a normal distribution, nor does it imply that the model output changes linearly with the inputs. A nonlinear model can transform a relatively simple input distribution into a skewed, asymmetric, or even multimodal output distribution.

A prediction distribution is conditional on its specification. Change the uncertain inputs, their distribution, the model, the dosing regimen, the endpoint, or the target population, and the resulting distribution may change. It is therefore incomplete to report “the predicted distribution” without identifying these elements.

Assumptions and the Target of Prediction

The mapping is straightforward to define, but its interpretation depends on the question. A distribution over a fixed drug-specific parameter can represent uncertainty about that parameter. A distribution over body size or physiology can instead represent differences among people in a target population. These are conceptually different sources of spread, even though both may enter the same input vector in a calculation.

For a prediction about a randomly selected person from a population, the input specification needs to represent the relevant population characteristics. If the question concerns uncertainty about a prediction for one specified person, the person’s characteristics may be fixed while uncertain model parameters are varied. In either case, state which inputs vary and why. Do not describe uncertainty about a parameter as biological variation among people, or describe population variability as uncertainty about a single person’s fixed characteristics.

The model and endpoint also need explicit conditions. For example, oral-dose AUC depends on the dose, the definition of AUC (such as complete AUC after one dose), and the model’s assumptions about dose proportionality and clearance. A concentration prediction requires a specific time after dosing and a clear concentration definition. If the model is not appropriate for the stated regimen or target, propagating its inputs does not make the prediction appropriate.

Worked Example: Mapping Input Scenarios to Oral-Dose AUC

Hypothetical setup. Consider a linear model for complete AUC after a single oral dose:

\mathrm{AUC} = \frac{F D}{CL}

Here \(D\) is dose in mg, \(F\) is the fraction of dose reaching systemic circulation, and \(CL\) is clearance in L/h. Assume dose-proportional pharmacokinetics, constant clearance, and complete AUC over the entire concentration-time course. The dose is fixed at 100 mg. For illustration, the specified input distribution assigns probability 0.60 to scenario A and 0.40 to scenario B.

Scenario	Probability	\(F\)	\(CL\) (L/h)
A	0.60	0.80	4.0
B	0.40	0.60	5.0

For scenario A, substituting the values gives:

\mathrm{AUC}_A = \frac{0.80 \times 100\ \mathrm{mg}}{4.0\ \mathrm{L/h}} = 20.0\ \mathrm{mg\cdot h/L}

For scenario B:

\mathrm{AUC}_B = \frac{0.60 \times 100\ \mathrm{mg}}{5.0\ \mathrm{L/h}} = 12.0\ \mathrm{mg\cdot h/L}

Thus the induced prediction distribution places probability 0.60 at 20.0 mg·h/L and probability 0.40 at 12.0 mg·h/L. Its mean is:

E(Y) = 0.60(20.0) + 0.40(12.0) = 12.0 + 4.8 = 16.8\ \mathrm{mg\cdot h/L}

The mean is one summary of the distribution; it does not replace the two possible scenario predictions. These scenario weights are hypothetical and illustrate the mapping only. Their interpretation depends on what the scenarios and probabilities represent for the target in the actual analysis.

Concentration at a Fixed Time

AUC aggregates exposure over time. A concentration at a fixed time is a different endpoint and can respond differently to the same uncertain inputs. To see the mapping explicitly, consider a hypothetical one-compartment oral model with first-order absorption and elimination. Under this model, for \(k_a \ne k_e\):

C(t) = \frac{F D k_a}{V(k_a-k_e)} \left(e^{-k_e t}-e^{-k_a t}\right), \qquad k_e = \frac{CL}{V}

Here \(V\) is apparent volume in L, \(k_a\) and \(k_e\) are rate constants in h\(^{-1}\), and \(C(t)\) is concentration in mg/L when \(D\) is in mg. This formula assumes the stated one-compartment structure, first-order absorption and elimination, and no additional processes such as lag time. It is used here to illustrate a mapping, not to establish that this structure is suitable for every drug.

Worked Example: Mapping Scenarios to Concentration at Two Hours

Hypothetical setup. The dose is 100 mg and the endpoint is concentration at \(t=2\) h. Suppose the input distribution assigns probability 0.70 to scenario A and 0.30 to scenario B. The model parameters are:

Scenario	Probability	\(F\)	\(k_a\) (h\(^{-1}\))	\(V\) (L)	\(CL\) (L/h)
A	0.70	0.80	1.0	20	4.0
B	0.30	0.60	1.2	25	5.0

For scenario A, \(k_e=4.0/20=0.20\) h\(^{-1}\). Substitution into the concentration equation gives:

C_A(2) = \frac{0.80(100)(1.0)}{20(1.0-0.20)} \left(e^{-0.20(2)}-e^{-1.0(2)}\right)

The prefactor is \(80/16=5.00\) mg/L. Using \(e^{-0.4}\approx0.67032\) and \(e^{-2}\approx0.13534\):

C_A(2) = 5.00(0.67032-0.13534) = 5.00(0.53498) = 2.6749\ \mathrm{mg/L}

For scenario B, \(k_e=5.0/25=0.20\) h\(^{-1}\). The prefactor is \(0.60(100)(1.2)/[25(1.2-0.20)]=72/25=2.88\) mg/L. Since \(e^{-2.4}\approx0.09072\):

C_B(2) = 2.88(0.67032-0.09072) = 2.88(0.57960) = 1.6692\ \mathrm{mg/L}

The induced distribution has probability 0.70 at 2.6749 mg/L and probability 0.30 at 1.6692 mg/L. Its mean is \(0.70(2.6749)+0.30(1.6692)=2.3732\) mg/L. This is the mean of the specified scenario distribution at two hours; it is not a claim that every person, or even a typical person, has that concentration.

Target-Population Weighting

A target population is not a label added after the calculation. It helps determine which input values and weights belong in the prediction. For example, if the population contains subgroups with different proportions, a prediction for a randomly selected member should reflect those proportions. Equal weighting of subgroups would answer a different question unless the subgroups are equally prevalent in the target or equal weighting is explicitly intended.

Worked Example: Changing the Target Population Changes the Prediction

Hypothetical setup. Suppose a PBPK model predicts fixed-time concentrations of 3.0 mg/L for subgroup A and 1.0 mg/L for subgroup B under the same regimen. Assume these are fixed subgroup predictions for this illustration. Target population 1 is 70% subgroup A and 30% subgroup B. Target population 2 is 40% subgroup A and 60% subgroup B.

For a randomly selected person from target population 1, the induced distribution has probability 0.70 at 3.0 mg/L and 0.30 at 1.0 mg/L. Its mean is:

E(Y\mid T_1) = 0.70(3.0) + 0.30(1.0) = 2.4\ \mathrm{mg/L}

For target population 2, the subgroup weights differ, so:

E(Y\mid T_2) = 0.40(3.0) + 0.60(1.0) = 1.8\ \mathrm{mg/L}

The model predictions within each subgroup have not changed; the population-level distribution has changed because the target composition changed. If there were additional within-subgroup variation or parameter uncertainty, each subgroup could have its own prediction distribution, and the target distribution would combine those subgroup distributions using the target’s weights.

Justification and Interpretation

The induced distribution follows directly from the probability assigned to input values. For a discrete input distribution with possible values \(\mathbf{x}_1,\ldots,\mathbf{x}_m\) and weights \(w_1,\ldots,w_m\), where \(w_i\ge0\) and \(\sum_i w_i=1\), each input value produces \(y_i=f(\mathbf{x}_i)\). The probability of an output value is the sum of the weights of all inputs that map to it:

P(Y=y) = \sum_{i:\, f(\mathbf{x}_i)=y} w_i

This also explains why distinct input combinations can yield the same output, as can happen when an endpoint depends on a ratio of inputs. For a continuous input distribution, the same event-based definition applies: \(P(Y\le y)\) is the probability assigned to all input vectors satisfying \(f(\mathbf{X})\le y\). The model need not have a convenient algebraic inverse for this definition to hold.

A single plug-in prediction evaluates the model at one chosen input vector. It can be useful as a reference calculation, but it does not show the spread or shape induced by an input distribution. Conversely, a prediction distribution does not by itself establish that its inputs are credible, that the model is accurate, or that the target is appropriate. The distribution inherits the assumptions and limitations of its ingredients.

Diagnostics, Failure Modes, and Limitations

Before interpreting a prediction distribution, check the mapping and its scope. The following questions help expose common failures:

Are the inputs defined for the stated question? Record units, fixed values, uncertain values, and the reason each value is included. An input distribution tied to a different population or regimen may not answer the current question.
Does every input combination have a valid model interpretation? Implausible combinations, invalid parameter values, or solver failures can create outputs that are computationally generated but scientifically meaningless.
Is the endpoint precisely defined? AUC requires a stated dose and integration definition; concentration requires a time and concentration measure. Changing the endpoint changes the mapping.
Are the weights meaningful for the target? In a scenario calculation, weights must sum to one and have a stated interpretation. Equal weights are not automatically representative of a population.
Does the result behave plausibly under simple checks? Recalculate selected inputs by hand, check units, and verify that the model output corresponds to the intended regimen. For the AUC example, increasing \(F\) increases \(FD/CL\), while increasing \(CL\) decreases it.

A frequent mistake is to treat a wide range of input values as if it were already a prediction distribution. A range alone specifies neither the probability of values within it nor their joint occurrence. Another mistake is to interpret an output distribution as a distribution of observed outcomes without checking what sources of variability and uncertainty it includes. The correction is to state the input distribution and target explicitly, then describe the output as conditional on that specification.

Comparisons with alternatives should preserve the question being answered. A plug-in prediction gives one model output for one chosen input vector. A scenario table shows outputs for selected cases, but it is a probability distribution only if its weights are justified and interpreted as such. A local approximation around one input vector may summarize small changes near that point, but it can miss nonlinear behavior or more distant input combinations. None of these choices is universally superior: the appropriate representation depends on the purpose, available information, model behavior, and intended target.

The prediction distribution also does not automatically account for model structural uncertainty, measurement error, or all individual differences. Those sources must be represented deliberately if the question requires them. A highly detailed numerical distribution can still be misleading if an important input, population feature, or model limitation is omitted. The distribution is a structured summary of the specified calculation, not a guarantee that the calculation captures reality.

Key takeaway: A PBPK prediction distribution is defined by the joint input distribution, the model mapping, the regimen, the endpoint, and the target. Interpret its spread only in light of which sources of uncertainty or population variability were actually represented.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A model maps uncertain inputs \(\mathbf{X}\) to AUC \(Y=f(\mathbf{X})\). What additional information is needed to interpret the resulting distribution as a prediction for a target population?
For a 100 mg oral dose, two equally weighted hypothetical scenarios have \((F,CL)=(0.50,2.0\ \mathrm{L/h})\) and \((0.80,4.0\ \mathrm{L/h})\). Using \(\mathrm{AUC}=FD/CL\), calculate each AUC and the mean of the two-point prediction distribution.
Why can two target populations produce different prediction distributions even if the PBPK model and subgroup-specific predictions are unchanged?
Explain why a range of possible clearance values alone is insufficient to define a probabilistic prediction distribution.
Is a distribution over a fixed drug-specific parameter necessarily the same thing as variability among people? Explain.

Solutions. (1) State the target and its relevant population characteristics, the input distribution and what sources its variation represents, the model, regimen, and exact endpoint. (2) The first AUC is \(0.50(100)/2.0=25.0\) mg·h/L; the second is \(0.80(100)/4.0=20.0\) mg·h/L. With equal weights, the mean is \(0.50(25.0)+0.50(20.0)=22.5\) mg·h/L. (3) Different subgroup proportions change the weights assigned to subgroup predictions, changing the population-level distribution. (4) A range gives possible values but not probabilities or weights, so it does not determine output probabilities. (5) No. Uncertainty about a fixed parameter and biological differences among people are distinct concepts; the analysis must say which source the distribution represents.