From Model Choice to a Worked Fit
Choosing between full target-mediated drug disposition (TMDD), a quasi-steady-state (QSS) approximation, and a Michaelis–Menten (MM) reduction is not only a theoretical exercise. A practical analysis must fit a model to observations, check what it predicts across dose and time, and determine whether the simplification changes the scientific conclusions. This tutorial follows one hypothetical simulated analysis through those steps.
The example uses intravenous bolus doses and measurements of free drug concentration \(C\) and free target concentration \(R\). The full model represents free drug, free target, and drug–target complex \(B\). Its equations below use concentration-equivalent states; rates and volumes are defined on a consistent scale. The initial free-target concentration is at baseline equilibrium, and the initial complex is zero. For a bolus, the initial drug concentration is the dose divided by the applicable volume.
The QSS model replaces the complex differential equation with a quasi-steady relationship \(B\approx CR/K_{\mathrm{SS}}\), where \(K_{\mathrm{SS}}=(k_{\mathrm{off}}+k_{\mathrm{int}})/k_{\mathrm{on}}\). The MM reduction represents the saturable drug-loss pathway directly, commonly as a term \(V_{\max}C/(K_m+C)\) in the drug equation. That reduction can describe drug concentrations, but it does not itself supply a dynamic free-target prediction. Consequently, a drug-only MM fit cannot be treated as a joint fit to drug and target outcomes.
The Simulated Data and Illustrative Fit
Suppose a hypothetical simulation includes 12 subjects receiving doses spanning low, intermediate, and high exposure. Drug and free-target measurements are collected at baseline, during the early distribution and binding period, and in the later decline. The table gives a small subset of the observations and fitted population predictions for one representative subject. It is a teaching summary, not a claim that these few points alone determine all NLME parameters.
| Time (h) | Observed free drug (nM) | Full TMDD prediction (nM) | QSS prediction (nM) | MM prediction (nM) | Observed free target (nM) | Full TMDD target prediction (nM) | QSS target prediction (nM) |
|---|---|---|---|---|---|---|---|
| 0.5 | 75.0 | 76.0 | 74.0 | 80.0 | — | — | — |
| 4 | 24.0 | 23.0 | 22.0 | 17.0 | 5.4 | 5.2 | 5.8 |
| 24 | 2.80 | 3.00 | 2.70 | 1.40 | 8.2 | 8.0 | 7.6 |
| Baseline | — | — | — | — | 10.0 | 10.0 | 10.0 |
For the full and QSS fits, assume the same drug and free-target observations are used with the same residual-error structure. The MM fit is to drug observations only; its prediction column is included for comparison on drug concentration, not as a target model. The parameter values and fit summaries that follow are hypothetical optimizer outputs from the larger simulated data set.
Worked Example: Interpreting Binding Parameters
Suppose the full TMDD population estimates are \(k_{\mathrm{on}}=0.120\ \mathrm{nM}^{-1}\mathrm{h}^{-1}\), \(k_{\mathrm{off}}=0.240\ \mathrm{h}^{-1}\), and \(k_{\mathrm{int}}=0.0600\ \mathrm{h}^{-1}\). Calculate the dissociation constant \(K_D\) and the QSS constant \(K_{\mathrm{SS}}\).
By definition, \(K_D=k_{\mathrm{off}}/k_{\mathrm{on}}\). Substitution gives \(K_D=0.240/0.120=2.00\ \mathrm{nM}\). The QSS constant includes both dissociation and complex removal: \(K_{\mathrm{SS}}=(k_{\mathrm{off}}+k_{\mathrm{int}})/k_{\mathrm{on}}=(0.240+0.0600)/0.120=2.50\ \mathrm{nM}\).
Thus the estimated QSS scale is \(0.50\ \mathrm{nM}\) larger than the dissociation constant. They are not interchangeable: complex internalization contributes to \(K_{\mathrm{SS}}\), but not to \(K_D\). These derived constants summarize fitted rates; they do not, by themselves, establish that the parameters are precisely estimated or identifiable from the available observations.
Worked Example: Checking Prediction Errors
At 0.5, 4, and 24 h, the observed drug concentrations are 75.0, 24.0, and 2.80 nM. The full-model predictions are 76.0, 23.0, and 3.00 nM. Calculate the root mean squared error (RMSE) for these three points.
Define the residual as observed minus predicted. The residuals are \(75.0-76.0=-1.00\ \mathrm{nM}\), \(24.0-23.0=1.00\ \mathrm{nM}\), and \(2.80-3.00=-0.200\ \mathrm{nM}\). Therefore,
This describes error for these three selected drug observations only. It is not the RMSE across all subjects and sampling times, nor does a small value at selected times demonstrate that the model is adequate across dose levels or for target measurements.
Comparing the Fitted Models
A fit should be assessed using both numerical summaries and the shape of predictions relative to the data. For example, the full-model predictions at 4 and 24 h are close to the listed free-target observations: 5.2 versus 5.4 nM and 8.0 versus 8.2 nM. The QSS predictions are 5.8 and 7.6 nM. Those differences are small in this subset, but the full data should be examined for systematic target residuals, especially near the early period when the approximation may be least accurate.
The drug predictions illustrate a different issue. At 4 h, the MM prediction is 17.0 nM, below the observed 24.0 nM, and at 24 h it is 1.40 nM, below the observed 2.80 nM. A pattern of negative observed-minus-predicted residuals at later times would indicate that the fitted MM model tends to predict excessive drug loss in that region. Such a pattern is more informative than a single overall error statistic because it can reveal dose- or time-dependent approximation failure.
Worked Example: Applying AIC to Comparable Fits
Suppose the full and QSS models are both fit to the same joint set of drug and target observations. The full model has \(-2\log L=214.6\) and \(k=10\) estimated parameters; QSS has \(-2\log L=220.1\) and \(k=8\). Calculate AIC for each model using \(\mathrm{AIC}=-2\log L+2k\).
For full TMDD, \(\mathrm{AIC}=214.6+2(10)=234.6\). For QSS, \(\mathrm{AIC}=220.1+2(8)=236.1\). The full model has the lower AIC by \(236.1-234.6=1.50\). Its improvement in fit is \(220.1-214.6=5.50\) units of \(-2\log L\), while its two additional parameters incur a penalty of \(2(10-8)=4.00\). The net AIC difference is therefore \(5.50-4.00=1.50\) in favor of full TMDD.
This is a modest AIC difference, not decisive evidence that full TMDD is universally superior. AIC compares relative expected information loss among the candidate models fitted to the same outcomes under comparable likelihood definitions. Because the MM model here is fit only to drug data, its AIC must not be ranked against these joint-likelihood AIC values. Drug-only predictive errors can be compared across all three models on common drug observations, but that comparison answers a narrower question.
Why the Comparison Requires Diagnostics
A lower objective function or AIC does not certify that a model has captured the mechanism or that its parameters are reliably estimated. Inspect observed-versus-predicted plots, residuals against time and predicted concentration, and separate summaries by dose and outcome type. For a joint fit, check drug and target residuals separately: strong drug predictions do not compensate for a systematic mismatch in free-target dynamics. Check that numerical integration is stable and that parameter estimates are not at implausible boundaries; numerical estimation and identifiability can affect interpretation even when the objective function converges.
The comparison also depends on the design. If measurements are sparse during rapid binding or absent at low doses, several parameter combinations may produce similar drug predictions. A reduced model can then fit well while its effective parameters do not identify the binding and target-turnover quantities of full TMDD. Conversely, if the data cover dose and time regimes where the approximation is adequate, QSS or MM may provide a useful, more parsimonious description for prediction. The choice should reflect the intended use, the measured outcomes, and whether the approximation performs acceptably over the relevant exposure range.
The main trade-off is therefore not simply “more parameters versus fewer.” Full TMDD can represent free target and complex processes, but may require richer data to estimate its parameters stably. QSS retains a target-related mechanistic structure while imposing an approximation on complex dynamics. MM can summarize saturable drug loss compactly, but may conceal the biological components that generated the apparent saturation. None of these models is guaranteed to extrapolate well outside the dose and time range used for assessment.
Check Your Understanding
Answer each question, then compare your reasoning with the short solutions below.
- Given \(k_{\mathrm{off}}=0.30\ \mathrm{h}^{-1}\), \(k_{\mathrm{int}}=0.10\ \mathrm{h}^{-1}\), and \(k_{\mathrm{on}}=0.20\ \mathrm{nM}^{-1}\mathrm{h}^{-1}\), calculate \(K_D\) and \(K_{\mathrm{SS}}\).
- Observed drug concentration is 10 nM and prediction is 12 nM. What is the observed-minus-predicted residual, and what does its sign mean?
- Two comparable models have \(-2\log L=100\), \(k=5\) and \(-2\log L=96\), \(k=8\). Which has lower AIC, and by how much?
- Why can’t a drug-only MM fit establish that free-target predictions are adequate?
Solutions. (1) \(K_D=0.30/0.20=1.50\ \mathrm{nM}\), while \(K_{\mathrm{SS}}=(0.30+0.10)/0.20=2.00\ \mathrm{nM}\). (2) The residual is \(10-12=-2\ \mathrm{nM}\); the model overpredicts the observed concentration by 2 nM. (3) The first AIC is \(100+2(5)=110\); the second is \(96+2(8)=112\). The first is lower by 2, despite its larger \(-2\log L\). (4) A drug-only MM equation predicts drug concentration, not free-target dynamics, so its drug fit contains no direct assessment of target predictions.