Diagnosing DDI Model Misspecification With Clinical Data

Diagnosing a Mismatch, Not Just Measuring One

A drug-drug interaction (DDI) model is useful only if its predictions are credible for the clinical conditions of interest. When observed victim-drug exposure differs from the prediction, the difference is a diagnostic signal—not, by itself, an explanation. The first task is to locate where the mismatch enters: in the perpetrator exposure that drives the interaction, in the predicted effect of that exposure on the victim drug, or in the time course of either drug.

Here, the perpetrator is the drug causing an interaction, and the victim is the affected substrate drug. A defensible comparison uses the same dose, regimen, sampling or exposure metric, and relevant clinical conditions in the prediction and the observed data. For example, a predicted unbound perpetrator concentration should not be compared directly with an observed total concentration and then treated as a model failure. A mismatch can also arise from dose timing, incomplete sampling, assay or measurement uncertainty, or differences in adherence.

A practical diagnosis compares three linked quantities: perpetrator concentrations over time, victim exposure expressed as an area-under-the-curve ratio, and the time course of the interaction. The comparison is most informative when these are examined in that order. If the perpetrator concentration prediction is already wrong, a victim AUCR mismatch may be an expected consequence rather than evidence that the victim fraction or inhibition parameter is wrong.

Set Up Comparable Predictions

For the victim, define the observed AUCR as the victim AUC with the perpetrator divided by the victim AUC without it, using matched conditions and the same AUC definition in both periods. The model-predicted AUCR is the corresponding ratio from the model. An AUCR above 1 indicates increased victim exposure; an AUCR below 1 indicates decreased exposure. These ratios describe exposure changes, not clinical outcomes by themselves.

Perpetrator comparison should likewise match the quantity that drives the model. Depending on the mechanism, that may be an unbound concentration, a concentration near a relevant site, or a concentration-time profile rather than a single total-plasma value. If the model uses a concentration-dependent inhibition term, using the wrong concentration metric can create an apparent error in the inhibition parameter.

Observed values have uncertainty, and predictions may have parameter and structural uncertainty. Therefore, a small difference does not automatically demonstrate misspecification. When uncertainty estimates are available, compare the discrepancy with their scale. For a simple approximate comparison of independent observed and predicted values, a standardized residual can be written as:

Here \(y\) is the quantity being compared and the two standard errors represent uncertainty in the observation and prediction. This calculation is only appropriate when the uncertainties are reasonably summarized on the same scale and their covariance can be ignored. In practice, inspect the data, prediction intervals, and residual patterns as well as any summary statistic. Repeated residuals in the same direction or a systematic time pattern can be more informative than one isolated difference.

A Simple Inhibition Model as a Diagnostic Tool

A stripped-down model helps clarify how a wrong victim fraction or inhibition strength can affect AUCR. Suppose a fraction \(f_m\) of victim clearance is through one inhibited pathway; the remaining fraction \(1-f_m\) is unaffected. Assume linear, time-invariant victim pharmacokinetics, unchanged non-inhibited clearance, and reversible inhibition represented by \(r=I_u/K_i\), where \(I_u\) is the relevant unbound perpetrator concentration and \(K_i\) is the inhibition constant. Under this simplified static approximation, the fraction of baseline clearance remaining is \((1-f_m)+f_m/(1+r)\). If exposure is inversely proportional to clearance, the AUCR is:

The expression follows by separating baseline clearance into an unaffected part and an inhibited part. The unaffected part remains \((1-f_m)\) of baseline clearance. Inhibition reduces the inhibited pathway to \(1/(1+r)\) of its baseline contribution, giving \(f_m/(1+r)\). Their sum is the remaining clearance fraction. With unchanged dose and bioavailability, AUC is inversely proportional to clearance, so the exposure ratio is the reciprocal of that fraction.

This is a diagnostic simplification, not a general DDI model. It does not represent time-dependent inhibition or induction, changing perpetrator concentrations, multiple interacting routes, or transporter effects. Its value here is that it shows why different parameter errors can produce similar AUCR mismatches. Even a well-measured AUCR may not identify which parameter is wrong.

Use AUCR to Probe Victim Fractions and Inhibition

If predicted and observed perpetrator concentrations agree on the relevant scale but victim AUCR does not, the discrepancy points more directly toward the relationship between perpetrator exposure and victim clearance. Candidate explanations include an incorrect victim fraction affected, a misspecified inhibition parameter, an omitted clearance route, or an effect not represented in the model. Agreement in concentration removes one possible explanation; it does not isolate one remaining parameter.

Read the Time Course, Not Only the Summary Ratio

An AUCR compresses an exposure profile into one number. It can conceal a mismatch in onset, peak effect, duration, or recovery. Comparing concentration and effect trajectories can help locate such patterns. For example, if perpetrator concentrations are captured accurately but victim concentrations rise more slowly than predicted, the model’s effect dynamics may be wrong even if the final AUC ratio is close. Conversely, an apparent delayed interaction may arise from missed doses or sampling times rather than from slow biological turnover.

Patterns That Guide the Investigation

The following patterns organize investigation without claiming that any one pattern proves a cause. Multiple mechanisms can produce similar discrepancies, and some mechanisms can act together.

For transporter-related predictions, an error in uptake or efflux contributions can make the victim response wrong even when the perpetrator concentration is well described. A model that omits a relevant transporter pathway may also attribute its effects to an enzyme fraction or inhibition parameter. The clinical discrepancy can reveal that the current representation is incomplete, but attributing it specifically to a transporter requires evidence that separates that pathway from alternatives.

Diagnostics, Alternatives, and Limits

Begin with data and comparability checks before revising model parameters. Confirm dose and sample times, units, assay basis, steady-state status where relevant, and whether concentration observations are total or unbound. Plot observed and predicted perpetrator concentrations together, then inspect victim concentration-time profiles and AUCRs. Look for consistent bias across subjects, dose levels, time points, or substrates rather than relying only on an overall mean.

Several approaches can help characterize a discrepancy, but they answer different questions. A plot or residual summary identifies where predictions differ from observations. Re-estimating parameters asks whether altered parameter values can improve fit, but parameters may be correlated and several combinations may fit equally well. Comparing alternative mechanistic structures can test whether a missing route or different time dependence is plausible, but a better numerical fit alone does not establish that structure as biologically true. New or more informative data may be needed to distinguish competing explanations.

• Check units and matching first. Comparing total observed concentration with an unbound prediction can create a spurious driver mismatch.
• Do not treat a fitted parameter as a diagnosis. A lower fitted victim fraction may compensate for an incorrect inhibition constant or omitted pathway.
• Do not infer adherence from one concentration. A single low value can reflect sampling time or individual variability; use dose records and the profile where available.
• Do not overinterpret an average AUCR. Opposing individual discrepancies or different time-course errors can cancel in a group summary.

The simple inhibition equation is not suitable for every clinical DDI. Time-dependent inhibition and induction, transporter-mediated effects, nonlinear pharmacokinetics, and changing disease or co-medication conditions may require richer models. Clinical observations also rarely identify every mechanism independently. The goal of diagnosis is to locate the mismatch, narrow plausible explanations, and state what evidence would distinguish them—not to force one cause from an underdetermined comparison.