Why a Plausible Ratio Can Still Be Misread
A PBPK drug-drug interaction (DDI) prediction can be numerically correct and still be used incorrectly. A label interpretation may overstate certainty by treating one simulated ratio as a guaranteed result, infer clinical harm directly from an exposure change, compare against the wrong reference regimen, or apply a decision threshold outside the context for which it was chosen. Each error changes the meaning of the prediction without changing its numerical value.
The earlier material on reading AUC and Cmax ratios for label decisions establishes that these are distinct exposure measures, and the material on separating exposure change from clinical consequence distinguishes pharmacokinetics from clinical outcomes. Here the focus is an audit: before using a predicted ratio in a labeling argument, identify exactly what it compares, what uncertainty surrounds it, and what decision context supports its interpretation.
Audit the Comparison Before Interpreting the Number
A ratio is meaningful only when its two conditions are specified. The denominator is not automatically “usual exposure,” nor is it necessarily exposure at the same dose as the DDI scenario. The reference could be the same victim-drug dose without the perpetrator, a different dose selected to represent a proposed label regimen, or another explicitly defined condition. These comparisons answer different questions.
For a conventional interaction comparison, the numerator and denominator should refer to the same victim-drug regimen and differ in the perpetrator condition. If the victim dose also changes, the resulting ratio combines the interaction with the dose change. That comparison may be useful for evaluating a proposed adjusted regimen, but it is not the interaction effect at a fixed victim dose. State the comparison in words before interpreting the ratio.
Likewise, identify the exposure endpoint. An AUC ratio describes a change in area under the concentration-time curve; a Cmax ratio describes a change in peak concentration. They need not be equal, and neither alone describes every feature of the concentration-time profile. A label argument should not silently substitute one endpoint for another or treat a result for one endpoint as a result for both.
Worked Example: Detecting a Mismatched Reference Regimen
Suppose a hypothetical PBPK model predicts a 24-hour AUC of 120 mg·h/L for a 100 mg victim-drug dose without an inhibitor, and 240 mg·h/L for the same 100 mg dose with the inhibitor. It also predicts an AUC of 120 mg·h/L for a 50 mg dose with the inhibitor. Calculate the interaction ratio at the original dose and the ratio for the proposed adjusted regimen versus the original reference.
Interaction at the same dose. The 100 mg dose is used in both scenarios, so the ratio isolates the modeled effect of adding the inhibitor under these conditions:
The model predicts twice the AUC with the inhibitor at 100 mg compared with 100 mg without it.
Adjusted regimen versus original reference. Comparing 50 mg with inhibitor to 100 mg without inhibitor gives:
This second ratio indicates equal predicted AUCs for those two specified regimens. It does not mean the inhibitor had no interaction effect: the same-dose comparison was 2.00. It also does not establish that the dose adjustment is clinically appropriate; that requires the relevant exposure-response and safety context.
A Point Prediction Is Not a Certain Individual Outcome
A PBPK point prediction summarizes a specified model calculation. It is conditional on model structure, parameter values, input data, and scenario assumptions. It is not a guarantee that every individual will have that ratio, and it does not by itself quantify how uncertain the prediction is. A point estimate may be accompanied by a predictive interval, sensitivity analysis, or other uncertainty analysis; those quantities answer different questions and should not be replaced by the point estimate.
A predictive interval, when available, describes a range of outcomes for a future simulated or observed individual under the assumptions of the prediction procedure. A confidence interval for an estimated quantity has a different interpretation. Neither interval removes uncertainty about whether the model assumptions adequately represent the intended population and conditions. If no uncertainty summary is reported, the appropriate conclusion is that the displayed result is a point prediction with uncertainty not quantified in that display—not that the result is certain.
Worked Example: Qualifying a Point Prediction Near a Decision Boundary
Consider a hypothetical decision rule defined for this example only: a predicted AUC ratio above 2.00 prompts further evaluation. A PBPK analysis reports a point prediction of 2.20, with a predictive interval from 1.40 to 3.10 for the stated population and regimen. What can be concluded?
The point prediction exceeds the example threshold because \(2.20>2.00\). The interval extends below and above it because \(1.40<2.00<3.10\). Thus, the point prediction falls on the “further evaluation” side of this particular rule, while the predictive interval indicates that predicted individual outcomes span both sides of the boundary.
It would be incorrect to say the interaction is certainly above the threshold, or that there is a 95% probability the true ratio lies in this interval unless that probability statement is justified by the interval’s construction and target. The careful conclusion is conditional: the point prediction exceeds the stated threshold, but the predictive range crosses it, so the decision is sensitive to outcome variability and should be interpreted with the model’s assumptions and the rule’s intended use in view.
Exposure Ratios Are Not Risk Ratios
An exposure ratio compares pharmacokinetic quantities. A risk ratio compares probabilities of an outcome between conditions. They have different numerators, denominators, and units of interpretation. An AUC ratio of 2.00 means the modeled AUC is twice as large in the numerator condition as in the denominator condition; it does not mean the probability of an adverse event doubles.
To infer a clinical risk change, one needs an appropriate relationship between exposure and the outcome, along with relevant baseline risk and other clinical information. That relationship may be nonlinear, may differ across patients, and may depend on which exposure measure is relevant. Without such evidence, a PBPK exposure prediction supports a statement about predicted exposure under specified conditions, not a numerical claim about adverse-event probability.
Worked Example: Why a Twofold AUC Does Not Imply a Twofold Risk
Suppose, hypothetically, a model predicts an AUC increase from 50 to 100 mg·h/L, giving:
Suppose a separate, validated clinical model—not the PBPK ratio itself—estimates adverse-event probabilities of 0.04 at 50 mg·h/L and 0.10 at 100 mg·h/L. The corresponding risk ratio in this hypothetical example is:
The exposure ratio is 2.00, while the risk ratio is 2.50. The figures differ because risk is determined by the specified exposure-response relationship and baseline risk, not by copying the exposure ratio. If the clinical model were unavailable, the risk ratio could not be calculated from these AUC values alone.
Thresholds Belong to Their Intended Context
A decision threshold is not a universal property of a drug or a ratio. Its meaning depends on the endpoint it applies to, the comparison and dosing conditions, the population, the decision being made, and the evidence or framework that supports its use. A cutoff for an AUC ratio cannot automatically be applied to a Cmax ratio. A rule used to categorize interaction magnitude does not, by itself, determine whether a particular patient should avoid a drug or change dose.
Before applying a threshold, verify that the predicted quantity matches the threshold’s endpoint and definition, and that the model scenario matches the conditions to which the rule is intended to apply. Do not select a threshold after seeing the result merely because it produces a preferred category. If the prediction is near a boundary, report the value and its uncertainty rather than presenting a categorical label as more decisive than the evidence permits.
Worked Example: Keeping AUC and Cmax Thresholds Separate
Assume a hypothetical analysis reports an AUC ratio of 2.10 and a Cmax ratio of 1.30. For this example alone, suppose the stated decision rule is defined for AUC and uses a threshold of 2.00. The rule is not defined for Cmax.
For AUC, \(2.10>2.00\), so the prediction exceeds the specified AUC threshold. For Cmax, the value 1.30 cannot be classified using that AUC rule because the rule’s endpoint does not match. It is valid to report both predicted ratios and to say that the AUC result exceeds the example threshold. It is not valid to claim that Cmax also exceeds that threshold, or to combine the ratios into an unspecified overall category.
A Practical Error-Check
A short audit before drafting a label-relevant interpretation can prevent the most common errors. It does not replace assessment of model credibility or clinical consequences; it makes the claim match the prediction that was actually generated.
Record whether the prediction is for AUC, Cmax, or another specified exposure measure, with its units before taking the ratio.
Specify victim dose, perpetrator condition, population, and other scenario conditions for each. Check whether the intended comparison changes only the factor of interest.
Distinguish a point prediction from a predictive interval or other uncertainty summary, and state when uncertainty has not been quantified.
Apply only a threshold defined for the same endpoint and context. Do not translate an exposure ratio into a risk ratio without an appropriate exposure-outcome model.
Common Mistake and Limits of Interpretation
Wrong reasoning: “The PBPK prediction is an AUC ratio of 2.0, so every patient’s exposure will double and their adverse-event risk will double; therefore any threshold of 2.0 supports the same label action.” This combines four unsupported steps: treating a point prediction as certain for each individual, equating exposure and risk ratios, ignoring the reference regimen, and applying a threshold without checking its context.
Correction: State the exact modeled comparison and endpoint; describe the result as a prediction under its assumptions; report available uncertainty; and limit the conclusion to what the evidence supports. A threshold can help organize a decision only when it is appropriate to the endpoint and use. Even then, it does not by itself establish clinical consequences or determine a dose action. PBPK predictions are also limited by model structure, parameter uncertainty, input-data quality, and the match between the simulated and intended population.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A predicted AUC ratio is 1.70 for a fixed dose with and without a perpetrator. What does the ratio say, and what does it not establish?
- A comparison uses 50 mg with an inhibitor in the numerator and 100 mg without it in the denominator. Why is this not a same-dose interaction ratio?
- A Cmax ratio is compared with a threshold defined for AUC. What is the interpretation problem?
- A point prediction is above a decision threshold, but its predictive interval crosses the threshold. Give a suitably qualified conclusion.
Solutions: (1) The modeled AUC in the numerator condition is 1.70 times that in the stated reference condition; this alone does not establish an individual’s response or an adverse-event risk ratio. (2) Both perpetrator condition and victim dose differ, so the ratio combines the interaction with a dose change. (3) The endpoint does not match the threshold; the AUC rule cannot classify the Cmax result. (4) State that the point prediction exceeds the threshold while the predictive interval spans values on both sides, making the result sensitive to predicted variability and the assumptions of the analysis.