Identifiability of Parent and Metabolite Potency

Why Separate Potencies Can Be Difficult to Estimate

A model can describe observed responses well while leaving the contributions of parent drug and active metabolite uncertain. This matters when the scientific question concerns which species drives the effect, or when a potency estimate will be used to predict a different dosing or exposure scenario. A good fit alone does not show that the individual potency parameters are determined by the data.

The central issue is whether the observations contain distinct information about each compound. If parent and metabolite concentrations rise and fall together, their predicted effects may be hard to separate. The problem can be exact, so that multiple parameter values produce identical predictions, or practical, so that predictions differ only slightly relative to measurement error. Uncertain metabolite formation and clearance can add further trade-offs by changing the concentration-time profile attributed to the metabolite.

Definition: A parameter is structurally identifiable if, under the stated model and ideal noiseless observations, distinct values of that parameter cannot produce the same observable predictions. It is practically identifiable if the available, noisy observations constrain it sufficiently for useful estimation. Structural non-identifiability cannot be fixed merely by collecting more observations of the same kind; practical imprecision may be reduced by more informative data.

A Model and the Source of Confounding

Let \(C_P(t)\) and \(C_M(t)\) denote parent and metabolite concentrations, and let \(R(t)\) be the response. A common additive model is

R(t)=E_0+E_{\max,P}\frac{C_P(t)}{EC_{50,P}+C_P(t)} +E_{\max,M}\frac{C_M(t)}{EC_{50,M}+C_M(t)}+\varepsilon(t).

Here \(E_0\) is baseline response, \(E_{\max,P}\) and \(E_{\max,M}\) are the maximum contributions of parent and metabolite, and \(EC_{50,P}\) and \(EC_{50,M}\) are the concentrations producing half of their respective maximum contributions. The residual \(\varepsilon(t)\) represents response error. This model assumes additive contributions on the chosen response scale, correct concentration-time profiles, and a specified error distribution. For the examples below, concentration is measured in mg/L and response change in response units. Unless stated otherwise, the maximum contributions and baseline are treated as known to focus on potency identifiability.

At concentrations much lower than the corresponding \(EC_{50}\) values, the approximation \(C/(EC_{50}+C)\approx C/EC_{50}\) gives

R(t)-E_0\approx \frac{E_{\max,P}}{EC_{50,P}}C_P(t) +\frac{E_{\max,M}}{EC_{50,M}}C_M(t).

In this range, the data primarily identify the ratios \(E_{\max}/EC_{50}\), not \(E_{\max}\) and \(EC_{50}\) independently. If the maximum effects are also estimated, low-concentration observations alone generally cannot separate maximum effect from potency. And if the concentration profiles are proportional, even the two ratios in this approximation may not be separately identified.

Key distinction: Similar parent and metabolite concentrations do not by themselves prove non-identifiability. The relevant question is whether their time courses, together with the response observations and model shape, provide distinct information about the parameters. Conversely, a narrow standard error from an unstable optimization does not establish identifiability.

Worked Examples

Worked Example: Exact Confounding Under Proportional Exposure

Suppose concentrations are all in the low-concentration range, the response baseline is known, and the approximate response model is \(R-E_0=a_P C_P+a_M C_M\). The coefficients \(a_P\) and \(a_M\) have units of response units per mg/L. Hypothetical observations have \(C_M=0.5C_P\), with \(C_P\) values of 1, 2, and 4 mg/L and response changes of 3, 6, and 12 units.

Substituting \(C_M=0.5C_P\) gives

$$ R-E_0=(a_P+0.5a_M)C_P. $$

The observation at \(C_P=1\) requires \(a_P+0.5a_M=3\). The other observations impose the same condition: \(6/2=3\) and \(12/4=3\). Thus the data identify the combination \(a_P+0.5a_M\), but not either coefficient separately. For example, \(a_P=2,\ a_M=2\) and \(a_P=1,\ a_M=4\) both give \(a_P+0.5a_M=3\), and therefore predict every listed response exactly.

The failure is structural for this design and model: no reduction in response measurement error would distinguish these coefficient pairs. In the nonlinear model, proportional exposures can also cause severe confounding, although curvature or observations spanning a wider concentration range may sometimes provide additional information. This example concerns the low-concentration approximation, so it does not establish that every nonlinear model with proportional profiles has exactly the same confounding.

Worked Example: A Profile Reveals Practical Confounding

Now suppose four hypothetical observations follow the same low-concentration model, with \(C_P=(1,2,3,4)\) mg/L and \(C_M=(1.0,1.9,3.1,4.0)\) mg/L. The response changes are generated by \(a_P=2\) and \(a_M=1\), giving \(y=R-E_0=(3.0,5.9,9.1,12.0)\) response units. Assume independent normal residuals with known standard deviation \(\sigma=0.1\) response units. The profiles are not exactly proportional, so the coefficients are structurally distinguishable in this linear model, but the small differences between the concentration profiles may make them practically confounded.

To profile \(a_M\), fix it at a candidate value and choose \(a_P\) to minimize the residual sum of squares. The least-squares value is

\hat a_P(a_M)= \frac{\sum_i C_{P,i}(y_i-a_M C_{M,i})}{\sum_i C_{P,i}^2} =\frac{90.1-30.1a_M}{30}.

For example, fixing \(a_M=0\) gives \(\hat a_P=90.1/30=3.003\). The residual sum of squares at that point is approximately \(0.01967\). Fixing \(a_M=1\) gives \(\hat a_P=2.000\) and residual sum of squares \(0\), because these values generated the observations. Fixing \(a_M=3\) gives \(\hat a_P=(90.1-90.3)/30=-0.00667\) and residual sum of squares approximately \(0.07867\).

With known \(\sigma\), the increase in negative twice the log-likelihood relative to its minimum is the increase in residual sum of squares divided by \(\sigma^2\). The increases for \(a_M=0\) and \(a_M=3\) are therefore \(0.01967/0.01=1.967\) and \(0.07867/0.01=7.867\). The first candidate is not excluded by the usual approximate one-parameter 95% profile-likelihood cutoff of 3.84; the second is. This profile demonstrates substantial uncertainty about \(a_M\), despite a perfect fit at the generating values. It is a diagnostic, not a proof that a particular confidence interval has exact coverage.

Worked Example: Concentration Range Helps Identify Potency

Consider hypothetical isolated-exposure observations under the full Emax model, with baseline removed and \(E_{\max,P}=E_{\max,M}=100\) response units known. During parent-only exposure, the parent concentrations are 1 and 36 mg/L and the response changes are 20 and 90 units. During metabolite-only exposure, the metabolite concentrations are 2 and 54 mg/L and the response changes are 25 and 90 units. Assume the concentrations and responses are measured without error for this illustration and the additive model is correct.

For the parent, \(20=100(1)/(EC_{50,P}+1)\). Rearranging gives \(20EC_{50,P}+20=100\), hence \(EC_{50,P}=4\) mg/L. The second parent observation checks this value: \(100(36)/(4+36)=90\) response units. For the metabolite, \(25=100(2)/(EC_{50,M}+2)\), so \(25EC_{50,M}+50=200\), giving \(EC_{50,M}=6\) mg/L. The second metabolite observation gives \(100(54)/(6+54)=90\) response units.

Here the compound-specific exposure periods separate the effects, and the concentrations span below and above the half-maximal concentrations. The example shows why observations that reveal curvature can help estimate \(EC_{50}\) when \(E_{\max}\) is known. If \(E_{\max}\) were unknown, two observations per compound would not generally be enough to establish both its maximum effect and its potency reliably; the numerical solution depends on the stated assumptions.

Parameter Profiles and PK Trade-Offs

A parameter profile checks how well the model can fit the data when one parameter is held at a succession of fixed values. At each fixed value, all other parameters are re-estimated. For \(EC_{50,M}\), for example, a profile asks whether changes in \(EC_{50,M}\) can be offset by changes in \(E_{\max,M}\), parent potency, baseline, or PK parameters while preserving a similar fit. A two-parameter profile can reveal a ridge where pairs of values fit nearly equally well.

\ell_{\mathrm{prof}}(\theta)=\max_{\eta}\ell(\theta,\eta),

where \(\theta\) is the parameter being profiled, \(\eta\) denotes all nuisance parameters, and \(\ell\) is the log-likelihood. For regular models with adequate information and an interior parameter value, a one-parameter 95% profile-likelihood interval is often defined by \(2[\ell_{\max}-\ell_{\mathrm{prof}}(\theta)]\leq 3.84\). Boundary values, weak identification, or small samples can undermine this approximation, so the cutoff should not be treated as universal.

Formation and clearance can create a related trade-off. In a joint model, a metabolite concentration-time course depends on its formation from parent and its own elimination. With sparse metabolite concentration measurements, a faster formation rate paired with faster clearance may yield predictions close to those from slower rates over the observed times. If predicted rather than measured metabolite concentrations drive the response model, that uncertainty can propagate into \(EC_{50,M}\): different formation, clearance, and potency combinations may produce similar responses. A response profile alone may hide this unless the relevant PK parameters are also allowed to vary or are profiled jointly.

Check the model and information source.
State which concentrations are measured and which are predicted, the response model, and which parameters are fixed or estimated. A potency estimate conditional on a fixed PK model does not describe uncertainty in that PK model.

Inspect parameter profiles.
Profile each potency over a range that includes values below and above the estimate. Re-optimize nuisance parameters at every point; a curve obtained by changing one parameter while freezing all others can exaggerate the information available.

Inspect joint trade-offs and predictions.
Look for ridges among parent potency, metabolite potency, formation, and clearance. Compare fitted predictions across the observed time course, not only the objective function at its minimum.

Report what the data support.
If a profile is broad or nearly flat, report the supported combinations or uncertainty rather than presenting an unstable point estimate as a precise compound-specific potency.

Diagnostics, Alternatives, and Limitations

Warning signs include a nearly flat profile, a long diagonal ridge in a two-parameter profile, large changes in potency estimates when a small number of observations are removed, and markedly different parameter estimates that produce nearly identical response predictions. A numerical optimizer may still return a single estimate in these situations. Convergence, a small residual sum of squares, or a Hessian-based standard error does not by itself establish that the parameters are identifiable.

One alternative is to estimate an effective combined contribution or a potency-normalized active-moiety measure when the data support that estimand but not separate potencies. This can be more honest for prediction within the observed exposure range, but it does not answer which species is responsible. Another option is to fix a PK or pharmacodynamic parameter using external evidence. Such constraints can stabilize estimation, but the resulting potency is conditional on the fixed value and should be accompanied by sensitivity analyses. Bayesian priors can similarly regularize weakly informed parameters; they add assumptions and information rather than making the observations intrinsically more informative.

Profile results also depend on model adequacy. A flat profile may reflect genuine confounding, but misspecified concentration predictions, an unsuitable response scale, or an incorrect error model can distort the profile as well. Conversely, an apparently sharp profile under a rigid model can be misleading if the model excludes plausible alternatives. The conclusions apply to the specified model and observed exposure conditions, not automatically to other doses, populations, or response mechanisms.

Common pitfall: “The optimizer returned different parent and metabolite EC50 estimates, so the data identify both potencies.” A successful numerical optimization only finds a solution under the chosen model and constraints. Check whether alternative potency and PK values yield a similar likelihood and similar predictions; if they do, the separate estimates are not reliably determined by these data.

Key takeaway: Separate potency estimates require observations that distinguish the parent and metabolite contributions, not merely a model that assigns them separate parameters. Use parameter profiles to expose trade-offs, and state clearly when only a combined effect is supported.

Check Your Understanding

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

If \(C_M(t)=0.25C_P(t)\) at every observation in the low-concentration linear model, which combination of effect coefficients is identified?
Why can low-concentration data have difficulty separating \(E_{\max,M}\) from \(EC_{50,M}\)?
In a profile-likelihood analysis, why are nuisance parameters re-estimated at each fixed potency value?
What does a broad, nearly flat potency profile imply, and what does it not prove?
How can uncertainty in metabolite formation and clearance affect the estimated metabolite potency?

Solutions. (1) The prediction is \((a_P+0.25a_M)C_P\), so the data identify \(a_P+0.25a_M\), not the two coefficients separately. (2) When concentrations are small relative to \(EC_{50,M}\), the metabolite term is approximately \((E_{\max,M}/EC_{50,M})C_M\); the observations mainly inform the ratio, not its numerator and denominator individually. (3) Re-optimization allows the model to make its best fit at each potency value, showing whether other parameters can compensate for that value. (4) It indicates weak information about the potency under the specified model and data. It does not, by itself, prove structural non-identifiability or identify the cause of the weak information. (5) Different formation and clearance values can imply different metabolite concentration profiles; if those predictions are uncertain, potency may trade off against them to produce similar responses.