Worked TMDD Analysis of Dose-Escalation PK/PD Data

Why Analyze Dose Escalation Mechanistically?

A dose-escalation study can reveal that exposure does not increase in direct proportion to dose, and that target or biomarker measurements change alongside drug concentrations. A mechanistic target-mediated drug disposition (TMDD) model offers one explanation to test: drug binds target, forms a complex, and the complex is removed or internalized. The value of the analysis is not simply that a nonlinear curve can be fitted. It is that the fitted states and rates make linked predictions for free drug, target, and response that can be checked against measurements.

The worked analysis below uses hypothetical data and a one-compartment model. Its conclusions are conditional on that model and on the assumed study design. In particular, a good fit to a few dose-level summaries is not enough to establish a unique mechanism or to identify every binding and turnover rate.

These equations assume a well-mixed compartment, mass-action binding, constant rate parameters, and the stated linear loss processes. They omit distribution into additional compartments, time-varying target synthesis, and between-subject variation. Those omissions are assumptions to evaluate, not facts guaranteed by the data. If infusion is stopped or a dose is given differently, the input term must represent that actual administration; the constant \(q\) form applies only during constant infusion.

Start With the Measurements and the Observation Model

Before fitting, specify what each assay measures. A free-drug assay is linked to \(C\); a total-target assay may measure \(R+B\), depending on assay behavior; and a complex assay is linked to \(B\). An assay that detects only free target should instead be linked to \(R\). These are different observation models, and substituting one for another can make the inferred mechanism misleading.

For example, suppose measurement errors are approximately additive and normally distributed on the original concentration scale. A possible observation model for free drug is \(C_{\mathrm{obs}}=C_{\mathrm{pred}}+\epsilon_C\), where \(\epsilon_C\sim N(0,\sigma_C^2)\); analogous equations can be specified for target and biomarker measurements. If errors are proportional, a log-scale error model may be more appropriate. The residual model should be chosen to reflect assay precision and the pattern of residuals, rather than selected only because it makes optimization convenient.

A joint fit estimates parameters by comparing model predictions with all relevant measurements at their recorded times. For independent normal residuals with known or estimated standard deviations \(s_i\), one common objective is the weighted sum of squared residuals, \(\sum_i[(y_i-\hat y_i)/s_i]^2\). Here \(y_i\) is an observation and \(\hat y_i\) its model prediction. Numerical integration is generally needed for time-course data. The fit should use the actual sampling times and doses; a table of dose-level averages cannot replace the time course when the rates governing transient behavior are of interest.

Steady-State Relationships as a Check

A useful analytical check is available under constant infusion at steady state. Define \(K_{\mathrm{ss}}=(k_{\mathrm{off}}+k_{\mathrm{int}})/k_{\mathrm{on}}\), a concentration. Setting \(dB/dt=0\) gives \(k_{\mathrm{on}}CR=(k_{\mathrm{off}}+k_{\mathrm{int}})B\), so \(B=CR/K_{\mathrm{ss}}\). Setting \(dR/dt=0\) and using that equality gives \(k_{\mathrm{syn}}=k_{\mathrm{deg}}R+k_{\mathrm{int}}B\). Finally, in the drug equation the binding and dissociation terms combine to \(-k_{\mathrm{int}}B\), yielding \(q=k_{\mathrm{el}}C+k_{\mathrm{int}}B\). These identities link predicted free drug, target, complex, and infusion rate.

The baseline target concentration \(R_0\) is the steady-state free target when no drug is present. The identities above do not say that target concentration remains at baseline during treatment. They also make an important identifiability point: steady-state binding is summarized by \(K_{\mathrm{ss}}\), not by separate estimates of \(k_{\mathrm{on}}\) and \(k_{\mathrm{off}}\). Time-course information and informative sampling are needed to distinguish processes that steady-state measurements combine.

A Worked Dose-Escalation Analysis

Consider a hypothetical study in which one representative subject receives three increasing constant-infusion rates, with sufficient time at each level for near-steady-state sampling. For illustration, suppose a joint fit to the fuller time-course data gives the parameter values in the preceding example and predicts \(C=0.100\), \(1.00\), and \(10.0\) nM at the three respective rates. The measurements below are hypothetical near-plateau observations; the predictions are the fitted model's values, not additional observed data. Assume the target assay measures total target \(R+B\).

For this illustration, define occupancy as the fraction of target in complex, \(B/(R+B)\). Link a normalized marker \(Y/Y_0\) to occupancy through \(Y/Y_0=1-I_{\max}B/(R+B)\), with \(I_{\max}=0.800\). This is a stipulated empirical PD link for the example, not an automatic consequence of TMDD. At the three fitted concentrations, the steady-state equations give \(R+B=1.00\) nM in each case, and \(B=C/(1+C)\) nM because \(K_{\mathrm{ss}}=1.00\) nM and \(R+B=1.00\) nM. The marker predictions can therefore be calculated directly.

The infusion-rate predictions also check against the same model. At \(C=0.100\) nM, \(q=0.0500(0.100)+0.100(0.09091)=0.01409\) nM/h. At \(C=1.00\) nM, \(q=0.0500(1.00)+0.100(0.500)=0.100\) nM/h. At \(C=10.0\) nM, \(q=0.0500(10.0)+0.100(0.90909)=0.5909\) nM/h. Thus the displayed infusion rates are consistent with the fitted steady-state relationships after rounding.

The observed free-drug values rise from 0.103 to 0.960 to 10.4 nM as infusion increases, and observed marker ratios decrease. That pattern alone is not a fit diagnostic: the total-target and time-course predictions matter too. In particular, the model predicts that free target decreases as complex occupancy increases, while total target remains at 1.00 nM for these specific parameter values. A systematic measured change in total target beyond assay variability would call for investigation rather than being dismissed because the free-drug curve looks plausible.

Parameter Uncertainty and Exposure-Response

A point estimate is not a measure of how well a parameter is determined. A local standard error can summarize uncertainty near an estimated optimum under the assumed model and residual distribution; profile likelihood or bootstrap intervals can be more informative when the likelihood is asymmetric or parameters are correlated. The following hypothetical uncertainty summary is assumed to come from a richer time-course dataset than the three rows shown above. The log-scale standard errors are specified so that the interval calculation is reproducible.

For the exposure-response interpretation, the example's fitted link predicts marker suppression of 7.27%, 40.0%, and 72.7% at the three concentrations, calculated as \(100(1-Y/Y_0)\). This is a model-based relationship between predicted target occupancy and a normalized marker. It is not by itself evidence that changing dose will cause the same marker change in another population, at another time, or for a clinical outcome. Establishing those claims requires appropriate study design, validation, and consideration of uncertainty in both the PK/target model and the PD link.

When fitting a real escalation dataset, uncertainty should be assessed for derived quantities as well as individual parameters. For example, predictions of free drug and occupancy at a proposed dose depend jointly on \(k_{\mathrm{el}}\), target turnover, binding, internalization, and the covariance among their estimates. Reporting a narrow interval for one rate does not guarantee precise exposure or response predictions.

Diagnostics, Alternatives, and Limitations

Inspect residuals against time, dose, and fitted values for each assay separately. Long runs of positive or negative residuals can indicate a missed time course feature; increasing residual spread with concentration can indicate a poor error model; and a dose-related bias in total-target predictions can signal a mismatch between the model and assay or biology. Also check that fitted states remain nonnegative, that predictions honor the actual dose history, and that solver tolerances are adequate. A fit that changes materially when numerical settings are tightened is not reliable.

Compare the mechanistic model with plausible alternatives using predictive checks and appropriate model-comparison tools. A model without explicit target binding may be a useful simpler comparator, but it cannot make the same free-target or complex predictions. A reduced quasi-steady-state approximation can be computationally convenient when binding is fast relative to the observed dynamics, but it imposes a different approximation and may not reproduce early transients. A direct empirical exposure-response curve may summarize the observed trend with fewer mechanistic assumptions, but it does not explain or predict the separate free-drug and target states. The preferred model is not automatically the most elaborate one; it should be supported by informative data and make adequate predictions for the intended use.

Dose escalation is especially vulnerable to confounding between dose level and time, cohort, or subject characteristics. If different dose levels are given to different subjects, a single typical-subject fit can conceal between-subject variation in target abundance or clearance. If escalation occurs within subjects, carryover and incomplete target recovery may matter. A steady-state analysis is inappropriate when sampling precedes equilibration, and sparse sampling may leave binding, internalization, and turnover weakly identified. These limitations should be stated when interpreting estimates or extrapolating beyond observed doses.