Translating Turnover Pharmacodynamic Models

Why Turnover Matters in Translation

A drug can change a biomarker without acting directly on the amount already present. It may alter the rate at which the biomarker is produced, while the biomarker continues to be removed through its usual turnover process. In that case, the observed response develops over time: a drug effect on production does not usually cause the biomarker to jump immediately to its new steady level.

When translating an animal indirect-response model to humans, the prediction depends not only on the translated drug effect, but also on the human biomarker baseline and turnover rate. Uncertainty in any of these inputs can matter, and their effects can depend on the prediction time. The calculations below treat the human-scale inputs as uncertain quantities informed by animal data and the translation assumptions. They do not assume that an animal parameter is automatically identical to its human counterpart.

This is one specific indirect-response structure: inhibition of production with first-order loss. A model in which the drug stimulates production or inhibits loss has a different equation and generally a different response profile. The interpretation of \(I\) also matters: it is the fractional inhibition of production at the stated exposure, not necessarily a direct percentage change in the measured biomarker.

From Animal Parameters to a Human Prediction

The translated prediction uses human-scale values for \(B\), \(k\), and \(I\) at the exposure of interest. The drug-effect term may depend on a concentration-effect relationship, so \(I\) can carry uncertainty from potency translation as well as from other effect parameters. Here, the effect at the relevant human exposure is represented directly by \(I\); translating potency itself is not the subject of this calculation.

The steady-state relation \(B=k_{\mathrm{in}}/k\) explains why parameterization requires care. If \(B\) and \(k\) are the uncertain inputs, then \(k_{\mathrm{in}}=Bk\) follows from them; it should not also be treated as an independent uncertain input without a model that defines their joint relationship. Keeping a coherent parameterization avoids counting the same uncertainty twice.

Propagating Parameter Uncertainty

A point prediction alone does not show how strongly the result depends on uncertain inputs. A standard local approximation is the first-order delta method. It is useful when parameter uncertainty is moderate, the prediction is smooth over the plausible parameter range, and a local linear approximation is adequate. It gives an approximate variance for the prediction; it does not make a poor structural model reliable.

The justification is a first-order Taylor expansion around the parameter means: \(g(\boldsymbol{\theta})\) is approximated by a constant plus \(\nabla g^{\mathsf T}(\boldsymbol{\theta}-E[\boldsymbol{\theta}])\). The variance of that linear combination is exactly \(\nabla g^{\mathsf T}\boldsymbol{\Sigma}\nabla g\). The approximation can be inaccurate when curvature is substantial across the plausible parameter range, for example when \(k\) is highly uncertain and the prediction time is long.

For \(Y=R(t)=B[1-I(1-e^{-kt})]\), the sensitivities are $$ \frac{\partial Y}{\partial B}=1-I(1-e^{-kt}),\qquad \frac{\partial Y}{\partial k}=-BIte^{-kt},\qquad \frac{\partial Y}{\partial I}=-B(1-e^{-kt}). $$

These derivatives are local changes in the prediction per unit change in each parameter. In particular, the turnover-rate sensitivity depends on time: it is zero at \(t=0\), and it approaches zero as \(t\) becomes very large. The prediction can therefore be most sensitive to \(k\) at intermediate times rather than at baseline or at the eventual steady state.

Covariance, Time, and Interpretation

The independent-input calculation is not appropriate when the translated parameters have material covariance. Earlier material on parameter covariance motivates retaining such dependence rather than varying each input as if the others were fixed. For the baseline, turnover rate, and effect, the approximate variance includes terms such as $$ 2\frac{\partial Y}{\partial k}\frac{\partial Y}{\partial I}\operatorname{Cov}(k,I). $$

At the example values, both \(\partial Y/\partial k\) and \(\partial Y/\partial I\) are negative. If \(k\) and \(I\) have positive covariance, their covariance term is positive and increases the prediction variance. Negative covariance instead reduces that term. This sign follows from the two sensitivities and the covariance; it should be checked rather than assumed. Baseline-effect or baseline-turnover covariance contributes analogous terms.

Assumptions, Diagnostics, and Alternatives

Before interpreting a translated prediction, check that the model's scientific and mathematical conditions are credible for the prediction target. The solution above assumes a stable pre-dose baseline at steady state, first-order loss, constant production inhibition over the interval, and no additional delay or feedback in the drug effect. It also assumes that the biomarker measured in the animal and the human prediction represent comparable quantities. If exposure changes substantially over time, \(I\) is not constant and this closed-form solution does not describe that exposure history.

Time-course data help distinguish the parameters. At \(t=0\), the prediction is \(B\), and its sensitivities to \(k\) and \(I\) are zero. At long times, the prediction approaches \(B(1-I)\), and its sensitivity to \(k\) approaches zero. Thus baseline observations inform \(B\), late observations help identify the magnitude of the inhibited steady-state change, and intermediate observations are especially informative about turnover. Sparse data limited to one time region may leave \(k\) and \(I\) difficult to distinguish.

Useful checks include comparing predicted and observed time courses, inspecting whether deviations are systematic over time, verifying units and parameter ranges, and checking whether the translated values yield plausible early and late responses. A poor fit at early times can indicate a missing delay or an incorrect exposure profile; a persistent late mismatch can point to an incorrect steady-state effect or baseline. A variance calculation cannot diagnose or repair these structural problems.

First-order propagation is efficient and shows which local sensitivities and covariances contribute to uncertainty. Its limitation is that it may misrepresent the prediction distribution when parameters are highly uncertain, constrained (for example, \(I\) must lie between 0 and 1), or the prediction is strongly nonlinear over their plausible values. A scenario analysis using coherent low, central, and high parameter sets can show how predictions change under selected assumptions, but those scenarios do not by themselves have probabilities or provide a calibrated uncertainty interval. A plug-in prediction using only mean parameters is simpler still, but conceals the uncertainty and may not equal the mean prediction when the model is nonlinear.

Common Pitfall

A common error is to translate the drug-effect magnitude and apply it immediately as a proportional change in the biomarker, for example predicting \(B(1-I)\) at every time after dosing. That expression is the eventual steady state under the stated model, not the time-course prediction. At finite time, the correct expression is \(B[1-I(1-e^{-kt})]\); the factor \(1-e^{-kt}\) accounts for turnover. A second error is to add parameter-variance contributions while omitting known covariances. The correction is to use the full covariance formula and to state clearly whether independence is an assumption or supported by the parameter estimates.