Structural Uncertainty in Cross-Species Models

Why Model Structure Matters in Translation

A cross-species prediction depends not only on the values assigned to parameters, but also on the equations that connect those parameters across species. For example, one model may assume clearance is proportional to body weight, while another assumes clearance is proportional to body weight raised to the 0.75 power. Both can be fitted to animal data, yet they may imply substantially different clearance in an adult human. That difference is structural uncertainty: uncertainty about which model form is an adequate representation of the translation process.

This is distinct from parameter uncertainty. If the exponent and model form are fixed, uncertainty in the fitted scaling coefficient is parameter uncertainty within that model. Structural uncertainty arises when scientifically plausible alternatives, such as different exponents or a physiological model instead of a simple allometric rule, remain under consideration. Sampling uncertain parameters from one model does not, by itself, account for uncertainty between model structures.

A useful comparison must state what is held constant and what changes. In the examples below, clearance is measured in L/h, body weight in kg, and species-level observations are treated as comparable measures of systemic clearance. The examples use hypothetical data; they are not results from a real study.

Fixed-Exponent Clearance Scaling

A common candidate structure is an allometric relationship between clearance \(CL_i\) and body weight \(W_i\) for species \(i\):

Here \(a\) is a scaling coefficient and \(b\) is a specified exponent. To fit the coefficient, take logarithms:

For the simple equal-weight log-scale fit used here, assume the species observations are independent, the same clearance definition applies across species, and the log-scale residuals have mean zero and a common variance. The candidate exponent \(b\) is fixed for each fit. These assumptions allow the fitted coefficient to be found by minimizing the sum of squared log residuals; they do not establish that the chosen exponent is biologically correct. With few species, the residual model is especially difficult to assess.

To see why, define \(z_i=\log(CL_i)-b\log(W_i)\) and \(c=\log(a)\). The log residual for species \(i\) is \(z_i-c\), so the sum of squared residuals is \(S(c)=\sum_i(z_i-c)^2\). Differentiating gives \(S'(c)=-2\sum_i(z_i-c)\). At the minimum, \(S'(c)=0\), hence \(c=\sum_i z_i/n\). Since \(a=\exp(c)\), exponentiating this mean gives the geometric mean of the ratios. The fitted intercept is therefore centered on the observed species in log space; it does not guarantee accurate prediction beyond them.

How Structural Choices Affect PK Predictions

Clearance is often an input to a PK prediction rather than the final quantity of interest. For example, under linear PK with complete systemic availability after an intravenous dose, total exposure is \(AUC = Dose/CL\). If the dose is held fixed, a higher predicted clearance implies lower predicted AUC. The assumptions matter: this relationship does not directly apply to nonlinear clearance, and an oral dose also requires accounting for bioavailability.

The same logic applies when clearance feeds into a PD prediction: a structural change in clearance can alter the concentration-time profile, which may in turn alter a concentration-driven effect. The size and direction of the PD consequence depend on the PK-PD relationship and the regimen. It is therefore preferable to propagate each candidate PK structure through the relevant prediction, rather than assume that a clearance difference has a uniform clinical meaning.

Comparing a Scaling Rule With a Physiological Structure

A simple allometric equation summarizes how clearance changes with body weight, but it does not explicitly represent the processes responsible for clearance. A physiological alternative may, for example, describe hepatic clearance using hepatic blood flow and intrinsic clearance. For the well-stirred liver model, let \(Q_h\) be hepatic blood flow and let \(f_uCL_{int}\) be unbound intrinsic clearance. The model is:

This structure assumes a well-mixed liver, with the stated flow and intrinsic clearance applying under the conditions being modeled. It concerns hepatic clearance; it is not total systemic clearance if important nonhepatic pathways are omitted. A comparison is meaningful only if the candidate models are being used for compatible quantities and conditions.

Diagnostics, Comparisons, and Limitations

Model comparison begins with the scientific assumptions, not simply with the numerical ranking of fit statistics. For each candidate structure, check that the animal data are comparable, the units are correct, and the fitted quantities have the same definition. Then inspect predictions and residuals on the scale used for fitting. A pattern in log residuals—for example, one species consistently above the fitted relationship and another below it—can indicate that a single power law is not describing the species pattern well. With only a few species, however, such patterns are fragile evidence, and a small residual sum of squares can arise from the particular observations.

It is useful to separate three questions:

• Fit: How closely does each structure describe the available animal observations, on an explicitly stated scale?
• Biological plausibility: Does the assumed scaling rule or physiological mechanism make sense for the drug and clearance pathway?
• Prediction: How different are the resulting human predictions, and are those differences consequential for the development decision?

A residual sum of squares can help describe fit when its scale and assumptions are stated, but it is not by itself a complete model-selection rule. Candidate structures with different numbers of parameters, different error assumptions, or different data targets cannot be ranked fairly by comparing unadjusted sums of squares alone. Likewise, the spread across candidate predictions is a useful scenario summary, but it is not automatically a confidence interval or prediction interval. It has no probability interpretation unless a justified probabilistic framework for model uncertainty has been specified.

A common pitfall is to select the model with the lowest animal-data residual sum of squares and report its human prediction as if structural uncertainty had been resolved. The correction is to report the assumptions, compare plausible alternatives, and show how the human prediction changes across them. If the predictions are decision-relevant, the difference should be made visible rather than hidden behind one selected model. A scenario range also has limitations: it only spans the structures considered, and it can miss important forms that were not included.

Other limitations include sparse species data, species-specific biology not represented by the model, and extrapolation from animals to a much larger human body size. A fixed-exponent comparison also treats each exponent as specified; uncertainty in the exponent itself is not captured by simply fitting separate fixed-exponent models. Nor does an apparent fit to animal data validate a human prediction. Structural comparison clarifies dependence on assumptions; it does not eliminate the need to evaluate whether those assumptions are appropriate.