Why the Same DDI Can Imply Different Parameters
A victim drug’s increase in exposure during inhibition depends on at least two quantities: the fraction of baseline clearance mediated by the inhibited pathway and the strength of inhibition at the enzyme site. A large interaction could arise because the pathway accounts for much of the victim’s clearance, because the inhibitor is potent at the relevant exposure, or from some combination of both. A fitted pair of values is not necessarily evidence that the observations have separately determined each value.
The question here is not how to obtain inhibitor potency from an enzyme assay. It is whether the DDI observations themselves contain enough information to distinguish victim fraction metabolized from inhibitor potency. The introductory material on estimating fraction metabolized from multiple evidence sources provides context for treating independent evidence as potentially informative; the focus here is the diagnostic behavior of DDI predictions and likelihoods.
This relationship assumes a baseline disposition that is already calibrated; a single inhibited pathway; unchanged clearance through all other pathways; a known, constant relevant \(I_u\); and linear victim pharmacokinetics so that AUC is inversely proportional to clearance. It also uses the specified inhibition factor \(1/(1+I_u/K_i)\) for the pathway’s remaining activity. The formula does not account for changing inhibitor concentrations over time, multiple inhibited pathways, or other perpetrator effects.
The expression becomes easier to interpret by defining \(x=I_u/K_i\), the exposure-to-potency ratio. Since \(I_u\) and \(K_i\) have the same concentration units, \(x\) is dimensionless. Define the inhibited fraction of baseline clearance, \(q=1-1/R\). Algebra gives:
Thus an observed ratio at a known \(I_u\) constrains a combination of \(f_m\) and \(K_i\). It does not, in general, determine both. In particular, if \(0<q<f_m\), rearranging gives \(x=q/(f_m-q)\), and hence:
Every admissible \(f_m\) greater than \(q\) therefore implies a corresponding \(K_i\) that reproduces the same ratio. This is structural non-identifiability from one exact observation: even without measurement error, there is a continuum of parameter pairs with identical predictions.
Worked Example: One Ratio, Two Parameter Pairs
Suppose a hypothetical study reports an AUC interaction ratio of \(R=1.50\) at a known unbound enzyme-site concentration \(I_u=2.00\ \mu\mathrm{M}\). Under the model, calculate \(q=1-1/R=1-1/1.50=1/3\). Consider two possible values of \(f_m\).
For \(f_m=0.500\), the required exposure-to-potency ratio is:
For \(f_m=0.800\), the same ratio requires:
Both parameter pairs predict \(R=1.50\) at \(I_u=2.00\ \mu\mathrm{M}\). The different \(K_i\) values are not a calculation error: each compensates for its different \(f_m\). The calculation retains \(q=1/3\) rather than substituting a rounded value into the rearranged expression.
What Additional Exposure Levels Can Reveal
Observations at different known inhibitor exposures can help separate the parameters because they test the shape of \(q\) as \(I_u\) changes. For fixed \(f_m\) and \(K_i\), \(q=f_m I_u/(K_i+I_u)\) rises with exposure and approaches \(f_m\) at high exposure. A single observation fixes one point on this curve; observations spanning a useful range can constrain both its scale and curvature.
The exposure range matters. At low exposures relative to \(K_i\), \(q\) is approximately proportional to \(I_u\), with slope \(f_m/K_i\). If all observations lie in that region, they may mostly determine the ratio \(f_m/K_i\), not the two quantities separately. At high exposures, \(q\) approaches \(f_m\), which can inform the fraction metabolized, but \(K_i\) may then be weakly determined because the curve changes little with further increases in exposure.
Worked Example: A Parameter Sweep Across Exposures
For a hypothetical parameter sweep, compare \((f_m,K_i)=(0.700,2.00\ \mu\mathrm{M})\) with \((0.800,18/7\ \mu\mathrm{M})\). Use known concentrations \(I_u=0.500,\ 2.00,\) and \(8.00\ \mu\mathrm{M}\). These pairs were chosen to give the same ratio at \(2.00\ \mu\mathrm{M}\); the question is whether their predictions separate at the other exposures.
For the first pair, \(q=f_m I_u/(K_i+I_u)\). At the three concentrations, the calculations give \(q=0.140,\ 0.350,\) and \(0.560\), respectively. Since \(R=1/(1-q)\), the predicted ratios are \(1.163,\ 1.538,\) and \(2.273\).
For the second pair, at \(I_u=0.500\ \mu\mathrm{M}\), \(q=0.800(0.500)/(18/7+0.500)=5.6/43=0.1302\), so \(R=1/(1-0.1302)=1.150\). At \(2.00\ \mu\mathrm{M}\), \(q=0.350\), giving \(R=1.538\). At \(8.00\ \mu\mathrm{M}\), \(q=0.800(8.00)/(18/7+8.00)=112/185=0.6054\), giving \(R=2.534\).
| \(I_u\) (\(\mu\mathrm{M}\)) | Ratio for \(f_m=0.700,\ K_i=2.00\ \mu\mathrm{M}\) | Ratio for \(f_m=0.800,\ K_i=18/7\ \mu\mathrm{M}\) |
|---|---|---|
| 0.500 | 1.163 | 1.150 |
| 2.00 | 1.538 | 1.538 |
| 8.00 | 2.273 | 2.534 |
The candidates agree at the middle exposure but diverge at the high exposure. Whether that difference is informative depends on the precision and reliability of the observations at that exposure. A sweep is a prediction comparison, not proof that the parameters can be estimated: if the available observations are noisy or cover only a narrow exposure range, both candidates may still be plausible.
Profile Likelihood as an Identifiability Diagnostic
A profile likelihood examines how well the data can be fitted when one parameter is fixed at a sequence of values and the other is re-optimized. For example, fix \(f_m\) at one candidate value, choose the \(K_i\) that gives the best fit, then repeat across candidate \(f_m\) values. If many values of \(f_m\) have nearly the same best fit after re-optimizing \(K_i\), the data provide weak information for separating them.
For clarity, suppose measured ratios \(R_j^{\mathrm{obs}}\) at known concentrations \(I_{u,j}\) have independent normal errors on the log scale, with known standard deviation \(\sigma_y\). Set \(y_j=\ln(R_j^{\mathrm{obs}})\) and \(\mu_j(f_m,K_i)=\ln R(f_m,K_i;I_{u,j})\). Apart from a constant, minus twice the log likelihood is:
The profile deviance at a fixed \(f_m\) is the smallest value of this expression over allowed \(K_i\). A shallow profile means that a broad range of \(f_m\) values fits almost as well as the best-fitting value. A narrow, well-defined minimum is more consistent with informative data, although it does not by itself validate the model assumptions. The likelihood model and its error scale must be appropriate for the observations; a profile is only as credible as the model used to construct it.
Worked Example: A Shallow Profile Despite Multiple Observations
Consider hypothetical log-ratio observations generated by the first pair in the parameter sweep: \(f_m=0.700,\ K_i=2.00\ \mu\mathrm{M}\), at \(I_u=0.500,\ 2.00,\) and \(8.00\ \mu\mathrm{M}\). Let the assumed standard deviation of each log-ratio measurement be \(\sigma_y=0.100\). The first pair fits these constructed observations exactly, so its residual sum of squares is zero.
Now fix \(f_m=0.800\) and use \(K_i=18/7\ \mu\mathrm{M}\). The middle prediction is unchanged. At the low exposure, the log-prediction difference from the constructed observation is approximately \(\ln(1.150/1.163)=-0.0113\). At the high exposure it is approximately \(\ln(2.534/2.273)=0.1089\). The approximate deviance increase relative to the exact fit is therefore:
This alternative pair is not an exact fit, but the increase in deviance is modest under the stated error model. The example shows why numerical optimization may return a particular pair even when the profile remains broad: the best-fitting point can be unique while nearby or distant combinations still fit nearly as well. The computed value is illustrative and depends on the hypothetical error assumption.
Practical Checks, Alternatives, and Limitations
Use parameter sweeps and profile likelihoods together. The sweep shows where predictions differ in the observed exposure range; the profile shows whether those differences are large enough, given the assumed variability, to distinguish parameter values. A practical workflow is:
Confirm that the baseline disposition is appropriate, that the modeled pathway is the intended inhibited pathway, and that \(I_u\) is known on the same concentration basis used by the model.
For each candidate \(f_m\), adjust \(K_i\) to match a reference observation, then compare predictions at every other observed exposure.
Fix one parameter across its plausible range, optimize the other, and examine whether the best fit changes appreciably.
If a broad ridge remains, report the trade-off or supported combinations rather than presenting one fitted pair as independently established.
An important diagnostic is the joint prediction pattern, not just the optimizer’s standard errors. A long, curved ridge in a joint likelihood or a nearly flat profile indicates strong confounding. Estimates pressed against a parameter bound can also signal that the data do not locate a clear optimum within the plausible range. Check whether alternative pairs give nearly identical ratios at all actual study exposures; predictions outside that range may diverge substantially without resolving the fit to the observed data.
A main alternative to estimating both quantities from DDI observations alone is to bring in independent information, such as a defensible estimate of \(f_m\) from other evidence, or observations spanning a broader and informative range of inhibitor exposures. Such information can reduce the trade-off, but its uncertainty and assumptions should be represented rather than treated as error-free when they are not. More observations at essentially the same exposure may improve precision for the ratio at that exposure without resolving the parameter pair.
The model also has boundaries. If the true interaction involves multiple affected pathways, changes in victim bioavailability, time-varying inhibitor exposure, or other perpetrator effects, a two-parameter clearance relationship may not describe the observations. Apparent parameter confounding can then be a symptom of model misspecification, not merely insufficient data. Conversely, a narrow profile under a misspecified model can give misleading confidence. The DDI observations alone do not establish that \(I_u\), the inhibition relationship, or the baseline clearance decomposition is correct.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- At one known inhibitor exposure, an observed ratio gives \(q=0.40\). What condition must \(f_m\) satisfy for a positive finite \(K_i\) under this model?
- Why might measurements at several low inhibitor exposures still fail to distinguish \(f_m\) and \(K_i\)?
- What does a broad, shallow profile likelihood for \(f_m\) indicate when \(K_i\) is re-optimized at each fixed value?
- Suppose two parameter pairs predict nearly identical ratios at every observed exposure but diverge at a much higher exposure not studied. Which prediction is directly supported by the data?
Solutions. 1. It must satisfy \(f_m>q=0.40\), with \(f_m\leq1\); otherwise the rearranged relationship does not yield a positive finite potency value. 2. When \(I_u\) is small relative to \(K_i\), \(q\) is approximately \(f_m I_u/K_i\), so the observations may mainly constrain \(f_m/K_i\). 3. Many values of \(f_m\) can be paired with an optimized \(K_i\) to fit nearly as well, indicating weak separate identification. 4. The ratios at the observed exposures are supported directly; the divergent high-exposure predictions are extrapolations, not an empirical distinction made by those data.