Combining PBPK and Population PK Evidence for a DDI Decision

Why Combine Mechanistic and Clinical Evidence?

A drug–drug interaction (DDI) decision often has two complementary sources of information. A physiologically based pharmacokinetic (PBPK) model can predict an interaction in a specific enzyme and dosing scenario, while a clinical population pharmacokinetic (popPK) analysis can estimate the interaction from observed concentration data while accounting for patient characteristics and sampling. The practical question is not simply whether the estimates differ. It is whether they address the same quantity, how uncertain each is, and what the combined evidence implies for coadministration.

The worked case below concerns a hypothetical victim drug whose exposure may increase when given with an enzyme inhibitor. Define the target as the population-level geometric mean ratio of victim-drug AUC with inhibitor to AUC without inhibitor, at specified doses and regimens. A ratio above 1 indicates increased exposure. The log of this ratio is convenient because multiplicative exposure effects become additive on the log scale.

Definition: Let \(R\) be the target AUC ratio for the specified population and dosing conditions, and let \(\theta=\log(R)\). The DDI estimand is the population-level log AUC ratio \(\theta\). AUC units cancel in the ratio, so both \(R\) and \(\theta\) are dimensionless. Uncertainty about \(\theta\) is not the same as between-patient variability in individual AUC.

For the evidence synthesis in this tutorial, the PBPK result is treated as a calibrated prior distribution for \(\theta\), and the clinical popPK result as a likelihood for \(\theta\). This interpretation requires both analyses to target the same population-level effect, endpoint, doses, inhibitor regimen, and relevant enzyme mechanism. The PBPK uncertainty must represent uncertainty about that target, including material parameter and model-discrepancy uncertainty; it must not be merely a prediction interval for individual patients.

Set Up the Evidence Before Combining It

Assume that the PBPK analysis supports a normal distribution for the target on the log scale, \(\theta\sim N(\mu_0,\sigma_0^2)\). Assume that the clinical popPK analysis provides an estimate \(y\) with standard error \(s\), and that, conditional on \(\theta\), \(y\sim N(\theta,s^2)\). For the simple update below, the clinical evidence is independent of the information used to construct the PBPK prior. The normal forms are working approximations; they should be checked against model uncertainty and the available evidence rather than adopted automatically.

The clinical standard error describes uncertainty in the estimated population interaction. It is not the standard deviation of individual patients’ AUC values. If the clinical model reports substantial interindividual variability, that variability belongs in the observation or population model; it should not be substituted for the standard error of the population interaction estimate.

Assumptions for the normal update: The PBPK prior and clinical likelihood concern the same \(\theta\); the clinical estimate is approximately unbiased with a usable standard error; the prior and clinical data are independent; and the reported uncertainty is on the log-ratio scale. If PBPK calibration used the same clinical observations, or if the endpoints or regimens differ, the simple independent update is not justified without further adjustment.

Deriving the Precision-Weighted Update

Under the stated assumptions, the prior density is proportional to \(\exp[-(\theta-\mu_0)^2/(2\sigma_0^2)]\), and the clinical likelihood is proportional to \(\exp[-(y-\theta)^2/(2s^2)]\). Multiplying these densities and collecting the terms involving \(\theta\) gives a normal posterior. The coefficient of \(\theta^2\) is the sum of the prior and clinical precisions, where precision is the inverse variance.

\begin{aligned} \frac{(\theta-\mu_0)^2}{\sigma_0^2} +\frac{(y-\theta)^2}{s^2} &= \left(\frac{1}{\sigma_0^2}+\frac{1}{s^2}\right) \left(\theta-\mu_{\mathrm{post}}\right)^2 +\text{constant},\\ \sigma_{\mathrm{post}}^2 &= \frac{1}{1/\sigma_0^2+1/s^2},\\ \mu_{\mathrm{post}} &= \sigma_{\mathrm{post}}^2 \left(\frac{\mu_0}{\sigma_0^2}+\frac{y}{s^2}\right). \end{aligned}

Thus, the posterior mean is a precision-weighted average: the estimate with smaller variance receives more weight. The posterior variance is smaller than either component variance when both precisions are positive. This reduction in uncertainty follows from the model assumptions; it is not guaranteed when evidence sources are dependent or do not estimate the same effect.

Worked Example: Interpreting the PBPK Prior

Suppose a hypothetical PBPK analysis predicts a population AUC ratio of 1.80. Its uncertainty is represented on the log scale by \(\mu_0=0.5878\) and \(\sigma_0=0.1335\). Check the central ratio and a 90% interval under the normal prior.

The central ratio corresponding to the log-scale mean is

\exp(\mu_0)=\exp(0.5878)\approx 1.800.

For a normal distribution, a central 90% interval is the mean plus or minus 1.645 standard deviations. On the log scale, the endpoints are

0.5878\pm(1.645)(0.1335) =0.5878\pm0.2196 =(0.3682,\;0.8074).

Exponentiating gives a ratio interval of approximately \((\exp(0.3682),\exp(0.8074))=(1.445,2.242)\). This interval describes uncertainty in the PBPK-based distribution for the population interaction under the stated model; it is not an interval containing 90% of individual patients’ AUC ratios.

Worked Example: Updating With Clinical Population PK Evidence

Suppose the hypothetical clinical popPK estimate of the log AUC ratio is \(y=0.5188\), corresponding to an estimated ratio of \(\exp(0.5188)\approx1.680\). Its standard error is \(s=0.0861\). Combine it with the PBPK prior from the preceding example. For transparent arithmetic, use the stated precisions 56.1 and 134.9, respectively.

The posterior mean is

\begin{aligned} \mu_{\mathrm{post}} &=\frac{(56.1)(0.5878)+(134.9)(0.5188)}{56.1+134.9}\\ &=\frac{32.97558+69.98612}{191.0}\\ &=\frac{102.96170}{191.0} \approx0.5391. \end{aligned}

The posterior variance and standard deviation are

\sigma_{\mathrm{post}}^2=\frac{1}{191.0}\approx0.005236, \qquad \sigma_{\mathrm{post}}\approx0.07236.

A central 90% posterior interval on the log scale is \(0.5391\pm(1.645)(0.07236)\), or approximately \((0.4200,0.6581)\). Exponentiating gives an interval of approximately \((1.52,1.93)\) for the AUC ratio. The posterior central ratio is \(\exp(0.5391)\approx1.715\). Under this model and prior, the clinical estimate shifts the central prediction down from the PBPK value while the additional evidence narrows uncertainty.

The interval is a Bayesian credible interval conditional on the prior, likelihood, and assumptions stated here. It is not a claim that there is a 90% long-run frequency of such intervals containing a fixed parameter, nor does it establish that the model is correctly specified.

Translate the Combined Evidence Into a Decision

A statistical estimate alone does not define a coadministration policy. The decision requires a clinically justified exposure threshold and an action rule set in context—for example, whether a dose adjustment, enhanced monitoring, or avoidance is warranted. In the next hypothetical calculation, suppose the development team has prespecified an AUC ratio of 1.5 as the point above which the increase is clinically important for this drug and has chosen to adjust the regimen if the posterior probability of exceeding that threshold is greater than 90%. This is a case-specific decision rule, not a universal DDI cutoff.

Worked Example: Probability of Exceeding a Clinical Threshold

Use the combined posterior \(N(0.5391,0.07236^2)\). The threshold \(R=1.5\) corresponds to \(\log(1.5)\approx0.4055\). Standardizing the threshold gives

z=\frac{0.4055-0.5391}{0.07236}\approx-1.846.

Therefore,

P(R>1.5\mid\text{evidence}) =P(\theta>0.4055\mid\text{evidence}) =1-\Phi(-1.846) \approx0.968.

Under the specified prior, likelihood, and decision rule, the posterior probability of an increase greater than 1.5-fold is about 96.8%, above the hypothetical 90% action threshold. The 90% posterior interval for \(R\), approximately 1.52 to 1.93, is also above 1.5. These calculations support the stipulated action to adjust the regimen or otherwise manage the interaction; they do not by themselves establish a particular safe dose.

For comparison, using the clinical estimate alone with a flat prior gives a normal posterior centered at 0.5188 with standard deviation 0.0861. Its threshold calculation is \(z=(0.4055-0.5188)/0.0861\approx-1.316\), so the probability of \(R>1.5\) is about 90.6%. The clinical evidence alone just exceeds this example’s action threshold, while the combined evidence gives a higher probability under the assumed independent-evidence model.

Diagnostics, Alternatives, and Limitations

Before relying on the update, check whether each source supports the shared estimand. A difference in AUC definition, sampling window, inhibitor dose, victim-drug regimen, enzyme contribution, or target population can make two apparently similar ratios non-comparable. Check also that the popPK standard error represents uncertainty in the population interaction estimate and that the PBPK distribution is calibrated for the same quantity. A very narrow prior can dominate the update; inspect the prior’s uncertainty and its sensitivity to plausible model discrepancy.

Independence deserves particular attention. If clinical data were used to tune PBPK parameters or to choose the PBPK model, those data are not wholly new evidence when reused in the likelihood. Treating them as independent can count information twice, making the posterior too precise. Where dependence is material, a joint model or an explicit covariance structure is needed; the independent precision-addition formula no longer applies as written.

The main alternatives answer different questions. A PBPK-only assessment is useful when clinical interaction data are absent, but its decision depends on the credibility and uncertainty of the mechanistic prediction. A clinical-only popPK assessment directly uses observed patient data, but may be imprecise or poorly transportable to an unobserved regimen. The combined update can improve precision when both sources are relevant and independently informative. It is not a license to average ratios arithmetically: the synthesis is on the log scale, and its weights come from variances.

The method also has limits. A normal approximation may be inadequate for skewed or multimodal uncertainty, and a single log-ratio may conceal clinically important subgroup variation or nonlinear exposure effects. The decision threshold must reflect the drug’s exposure-response and safety context. Even a well-supported exposure prediction does not prove a clinical outcome, and translating an AUC ratio into a dose change requires additional assumptions, such as dose proportionality and a sufficiently stable interaction across doses.

Common pitfall: “The PBPK and popPK ratios are close, so averaging them gives a reliable answer.” Closeness alone does not justify combination. First confirm a common estimand and independent evidence; then combine on the log scale using uncertainty. If dependence, endpoint mismatch, or poor calibration is present, a narrower combined interval can be misleading rather than reassuring.

Align the target.
Specify the population, victim drug, inhibitor regimen, AUC definition, and population-level log ratio.

Represent uncertainty.
Put the PBPK predictive distribution and clinical estimate with its standard error on the same log scale.

Check evidence dependence.
Confirm the clinical data were not already used in constructing or calibrating the PBPK prior.

Update and decide.
Calculate the posterior distribution, compare it with a clinically justified threshold, and apply the prespecified action rule.

Check Your Understanding

Answer each question, then compare with the solutions below.

Why is the interaction combined on the log-ratio scale rather than by averaging the AUC ratios?
In the worked case, what does the 90% interval from about 1.52 to 1.93 describe?
What is one reason the independent precision-weighted update could give an overly narrow posterior interval?
If the threshold ratio is 1.5, why does a posterior probability above 90% not establish that the combination is safe?

Solutions. (1) The ratio is multiplicative, while its logarithm is additive; the stated normal prior and likelihood are on the log scale, so their uncertainties combine through precision weighting there. (2) It is a 90% posterior credible interval for the population AUC ratio, conditional on the model, prior, clinical likelihood, and assumptions; it is not an interval for individual patient ratios. (3) The same clinical data may have contributed to PBPK calibration, so treating the prior and likelihood as independent would count some information twice. (4) The probability concerns exceeding the specified exposure threshold, not clinical safety itself. Safety also depends on the exposure-response relationship, the threshold’s justification, patient variability, and whether the proposed management is adequate.