Defining Translational Prediction Targets

Why Define the Prediction Target First?

A preclinical-to-human prediction is useful only when it answers a clearly stated question. “Predict human pharmacokinetics” is not yet a precise target: it does not specify which human population is meant, what dose or regimen applies, which PK quantity is required, or whether the result should describe a typical individual or variation across individuals. The same predicted clearance can imply different exposures at different doses, and a typical exposure does not directly state how many individuals might exceed an exposure threshold.

This tutorial defines prediction targets for human clearance, exposure, and pharmacodynamic (PD) response. The focus is on what quantity is to be predicted and how to describe it—not on a particular method for translating animal data or quantifying the uncertainty in that translation. The introductory material on first-in-human dose selection provides the decision context; here, the essential step is to make the human prediction question explicit before selecting or interpreting a model output.

Typical Individuals and Population Distributions

A population PK or PK/PD model can describe both a typical individual and a distribution of individual outcomes. In a common model representation, an individual parameter is written as a typical parameter multiplied by an individual-specific factor. For example, let \(CL_i\) be the clearance of individual \(i\), and let \(CL_{\mathrm{typ}}\) be clearance for a specified typical covariate profile. One possible model is:

Under this lognormal model, setting \(\eta_i=0\) gives the typical individual’s clearance, \(CL_{\mathrm{typ}}\). It is also the median clearance in the population described by this model. It is not generally the arithmetic population mean: that mean is \(CL_{\mathrm{typ}}\exp(\omega^2/2)\). A “typical individual” therefore needs a definition tied to the model and covariate profile; it should not be casually treated as an average person or as the population mean.

A prediction for the population distribution answers a different question. It can describe the proportion of individuals expected to fall above or below a threshold, or a selected quantile, conditional on the target population and model assumptions. Such a prediction is not interchangeable with a point prediction for the typical individual. Nor does the distribution itself identify which individual will have a particular value.

Specifying Clearance and Exposure Targets

A clearance target should identify the clearance definition and units, the human population or reference covariate profile, and whether the target is a typical value or a population summary. For example, “typical systemic clearance in litres per hour for a 70-kg reference adult” is more interpretable than “human clearance.” If clearance varies with covariates, the reference profile matters. If a population distribution is required, name the requested summary—such as the median, mean, or 90th percentile—and the population to which it applies.

An exposure target also requires the dose and regimen, route, and exposure metric. AUC and maximum concentration \(C_{\max}\) are not the same endpoint. AUC over one dosing interval is not necessarily the same quantity as AUC over a specified study period. Thus “human exposure” is incomplete unless the metric and its time interval are given. The target may be a typical-individual AUC at the proposed regimen, a population median AUC, or the probability that AUC exceeds a specified value.

For a fully bioavailable intravenous dose \(D\), under linear pharmacokinetics with constant clearance, total AUC over the complete disposition of that dose is \(D/CL\). This relation helps define a target; it does not make the dose or assumptions optional. With oral dosing, incomplete bioavailability changes the relation, and with nonlinear PK, exposure need not be proportional to dose. The intended population and summary still need to be specified in either case.

Specifying a Pharmacodynamic Target

A PD target must identify the response variable, its measurement scale and units, the exposure or dose condition, and—when relevant—the time point or interval. “Predict response” could mean a biomarker value at a specified time, a maximum effect, a change from baseline, or the probability of crossing a clinically or biologically defined threshold. These targets answer different questions.

The target must also state whether response is for a typical individual or a population summary. For a response threshold \(R^*\), one useful population target is \(P(R_i\geq R^*)\), the probability that an individual in the specified population has a response at least as large as \(R^*\), under specified exposure conditions. This probability is meaningful only after defining the population and the model for individual response. A typical response at a reference profile does not by itself provide this probability.

Choosing and Checking the Target

The target should follow from the decision being supported. If the question is what exposure is expected for a reference adult at a proposed dose, a typical-individual prediction may be appropriate. If the question is how many individuals might exceed a safety-relevant threshold, the target must describe the distribution or the relevant exceedance probability. If the question concerns a biomarker at a specified time, a time-specific response target is needed rather than an unspecified “PD effect.”

Before interpreting a result, check that its units and conditions match the target. Clearance has units of volume per time; AUC has units of concentration multiplied by time; a response may be measured in biomarker units or as a probability. Confirm that the route, dose, regimen, time interval, covariate profile, and population are all stated as needed. If the model’s output is for a reference individual but the decision concerns a population tail, the mismatch is a target-definition problem, even if the model calculation itself is correct.

A useful specification can be written in one sentence: “For [defined population and covariate conditions], predict [endpoint and units] under [dose, regimen, and time conditions], summarized as [typical value, mean, quantile, or probability].” This statement makes the intended quantity auditable and helps distinguish the prediction target from the particular model or data used to estimate it.

Common Pitfalls and Alternatives

A frequent mistake is to report the model’s typical value as though it were the population average. For a lognormal parameter model, the typical value is the median, while the arithmetic mean is larger when the between-individual variance is positive. Another mistake is to infer a population threshold probability from a typical response alone. A typical response of 50 units does not reveal the fraction above 40 units; that requires a distributional model and a defined threshold.

A point estimate and a distributional prediction are alternatives only when they answer the same decision question, which they often do not. A point estimate is compact and can be useful for a reference scenario. A mean summarizes expected magnitude on the original scale, while a median describes the central value in a way less affected by a long upper tail. A quantile or exceedance probability is more directly aligned with some threshold decisions. No summary is universally best: the target must be chosen before the summary is reported.

The resulting target is also conditional on the assumptions used to define it. For example, the exposure calculation above assumes linear PK and complete intravenous bioavailability; a different route or nonlinear relationship would require a different exposure calculation. A target stated without its conditions can appear more general than it is. Conversely, specifying a target does not validate the model or guarantee that the prediction is accurate; it ensures that the result is compared with the question it was meant to answer.