Diagnosing Approximation Failure Across Dose and Time

Why Approximation Error Must Be Checked Across Dose and Time

Quasi-steady-state (QSS), rapid-equilibrium (RE), and Michaelis–Menten (MM) reductions can make TMDD models easier to compute or interpret. Their predictions are not automatically interchangeable with those of the full target-mediated drug disposition (TMDD) system. A reduction replaces part of the binding and complex dynamics with an assumption. That assumption may be adequate during one dose range or time window and inadequate in another.

The central diagnostic is a matched prediction comparison: provide the full model and each reduced model with the same dose, sampling times, initial conditions, and parameter values, then compare their predicted drug and target profiles. This isolates approximation error from differences caused by refitting parameters, changing the regimen, or comparing different observations. The aim is not simply to find one summary score. The pattern across dose, time, and measured quantity can show when a reduction is failing and what profile is affected.

Models and Conditions Being Compared

Write \(C\) for free-drug concentration, \(R\) for free-target concentration, and \(RC\) for drug–target complex concentration. For a simple one-compartment TMDD system in concentration units, a representative full model is

Here \(u(t)\) is the drug input rate in concentration per unit time; \(k_{\mathrm{el}}\), \(k_{\mathrm{on}}\), \(k_{\mathrm{off}}\), \(k_{\mathrm{deg}}\), and \(k_{\mathrm{int}}\) are nonnegative rate constants; and \(k_{\mathrm{syn}}\) is the target synthesis rate. The model’s measured output must be defined consistently: a free-drug assay is compared with \(C\), while a total-drug assay may correspond to \(C+RC\), subject to the assay definition. A target assay may measure \(R\), total target \(R+RC\), or another specified quantity. Comparing different observables would not diagnose approximation error.

The QSS approximation imposes an algebraic relation for complex by treating its rate of change as negligible relative to the slower drug and target processes. Setting \(dRC/dt\) to zero in the complex equation gives \(RC\approx CR/K_{\mathrm{SS}}\), where \(K_{\mathrm{SS}}=(k_{\mathrm{off}}+k_{\mathrm{int}})/k_{\mathrm{on}}\). RE instead uses \(RC\approx CR/K_D\), where \(K_D=k_{\mathrm{off}}/k_{\mathrm{on}}\), and assumes binding and dissociation are sufficiently rapid to maintain equilibrium. The MM reduction describes drug disposition using a saturable elimination term such as \(V_{\max}C/(K_m+C)\); in its usual drug-only form, it does not predict a free-target profile.

These approximations require more than the formulas to be computationally convenient. QSS requires complex relaxation to be fast enough that the complex remains close to its algebraic relation as drug and target change. RE requires rapid binding and dissociation relative to the changing system and, for the equilibrium relation to represent the QSS relation, internalization must be small relative to dissociation. MM requires the saturable elimination representation to be appropriate for the drug profile of interest; its parameters and assumptions do not automatically preserve explicit target dynamics.

A Short Justification for Comparing QSS and RE Separately

The distinction between QSS and RE follows directly from their binding constants:

Thus \(K_{\mathrm{SS}}=K_D\) only when \(k_{\mathrm{int}}=0\), and the two constants are close when \(k_{\mathrm{int}}/k_{\mathrm{off}}\) is small. For example, if \(k_{\mathrm{int}}/k_{\mathrm{off}}=0.25\), then \(K_{\mathrm{SS}}=1.25K_D\). Even then, RE can fail if the binding process does not equilibrate quickly relative to the changing drug and target concentrations. Similar constants do not by themselves establish similar time courses.

Calculate Discrepancies on a Dose–Time Grid

The following hypothetical numerical benchmark illustrates the comparison. The entries are model predictions, not observations from a real study. Each model is evaluated under the same regimen and measurement definition; the full-model predictions are the reference for calculating discrepancy. Drug concentrations are in mg/L. For compactness, this grid shows free drug only; the next example considers target predictions.

Check Drug and Target Profiles Together

Drug concentration alone may not reveal whether a reduction reproduces target dynamics. A reduction can approximate free drug at selected samples while misrepresenting free target or total target. The model comparison must therefore include each observable that matters to the analysis. MM in its usual drug-only form has no target state to compare; “no prediction” is not a zero target prediction, and it cannot be scored as a target-profile match.

Practical Diagnostics and Failure Modes

Use a grid of doses and sample times that spans the intended application, rather than choosing only a dose or time where the models happen to agree. At each grid point, retain the signed discrepancy and the absolute discrepancy. The sign shows whether the approximation overpredicts or underpredicts; the absolute discrepancy describes the magnitude in the outcome’s units. Percentage discrepancies are useful when full-model predictions are not near zero, but can become unstable when the denominator is very small. For strictly positive predictions, a log ratio, \(\log(\widehat{Y}_{\mathrm{red}}/\widehat{Y}_{\mathrm{full}})\), is another scale-aware summary; it is undefined for nonpositive predictions.

A systematic sign change over time, a growing late-phase discrepancy, or errors concentrated at low or high doses is more informative than a single average. Also inspect target and drug outputs separately. If model predictions are displayed on a logarithmic axis, add a discrepancy summary on a clearly defined scale so that visual closeness is not mistaken for a small relative difference.

• Compare like with like. Match free drug to free drug, total drug to total drug, and the specified target measurement to that same observable. Confirm units, dose timing, initial conditions, and sampling times.
• Separate approximation error from refitting. First compare predictions using shared, corresponding parameter values. If each model is separately fitted, a changed prediction can reflect both approximation and parameter adjustment; that is a different comparison.
• Check the numerical reference. The full TMDD ODE solution should be numerically stable at the times and doses examined. If solver settings materially change its predictions, the reference itself is not reliable.
• Look beyond averages. Opposite signed errors can cancel in a mean. Report or plot errors by dose, time, and profile, not only as one pooled score.
• Do not score an absent state as a match. A drug-only MM reduction cannot establish target-profile adequacy without an explicitly defined target model.

How to Read the Comparison

Full TMDD is the mechanistic reference for this diagnostic when its state equations and numerical solution are appropriate. QSS retains a target-related algebraic relationship while simplifying complex dynamics; RE makes the stronger equilibrium assumption. MM provides a saturable drug-disposition representation, but its usual form does not preserve explicit target and complex profiles. These approaches therefore differ in both their assumptions and the quantities they can predict.

A discrepancy is evidence that the approximation does not reproduce the full-model prediction under the tested conditions; it is not, by itself, proof that the full model describes real data correctly. Conversely, close predictions over a limited grid do not prove agreement outside that grid or under a different regimen. The diagnostic is conditional on the chosen parameter values, outputs, and dose–time coverage. It does not replace assessment against observations, and it does not determine which model should ultimately be used.