Defining Translational PK/PD Estimands

Why Define the Estimand Before Translating

A preclinical result does not, by itself, answer a human prediction question. “The drug reduced the biomarker” is incomplete unless the species, population, dosing regimen, observation time, biomarker definition, and summary are specified. The same is true for pharmacokinetics: a clearance value or exposure summary is meaningful for translation only when its dose, route, time window, and population are clear.

An estimand is the precisely defined quantity an analysis aims to describe or predict. It is not the dataset, the statistical model, or the numerical estimate obtained from that dataset. In translation, distinguish the source-species estimand (what is being summarized in the preclinical setting) from the target human estimand (what quantity is sought under a specified human scenario). A model may connect them, but that connection is an assumption to justify, not part of the definition.

For example, “mean unbound plasma AUC from zero to 24 hours after a single 10 mg/kg oral dose in healthy adult humans” is more precise than “human exposure.” The precise statement identifies a population, regimen, matrix, interval, and population summary. If the intended use concerns patients with organ impairment or repeated dosing, that is a different estimand, even if the drug and endpoint are unchanged.

Four Quantities That Need Explicit Definitions

Clearance

Clearance is a volume of plasma or blood cleared of drug per unit time. For an individual after an intravenous dose under linear, time-invariant pharmacokinetics, systemic clearance can be estimated as dose divided by total plasma concentration AUC from zero to infinity. The relevant estimand must say whether it is an individual clearance, a population mean or median, or another summary, and whether it refers to plasma or blood.

This expression is not a universal recipe for every dataset. It relies on an intravenous dose, an adequately characterized AUC to infinity, and conditions under which the dose-to-AUC relationship supports a single clearance value. Nonlinear disposition, incomplete sampling, or a different route can require a different definition or estimation method. Clearance is not itself an exposure or a response: it is a PK quantity whose interpretation depends on the specified conditions.

Unbound exposure

Unbound exposure concerns the concentration not bound to plasma proteins in the stated matrix. If \(C_{tot}(t)\) is total concentration and \(f_u(t)\) is the unbound fraction at time \(t\), then unbound concentration is \(C_u(t)=f_u(t)C_{tot}(t)\). Its AUC over a specified interval is the integral of that concentration over time. If the unbound fraction is constant over the interval, it can be taken outside the integral; otherwise, multiplying total AUC by one unbound fraction may not represent the interval-specific unbound AUC.

An unbound-exposure estimand therefore names the matrix, interval, regimen, and population summary. Measurement conditions matter: a fraction measured under conditions that do not represent the relevant matrix or concentration range may not describe the unbound exposure of interest.

Target engagement

Target engagement is an operationally defined measure of drug interaction with, or modulation of, a specified biological target. It may be measured directly, such as by an occupancy assay, or indirectly using a validated target-linked biomarker. The estimand must identify the assay, target, tissue or matrix, time after dose, and summary. A percentage from one assay is not automatically interchangeable with a percentage from another assay or tissue.

For an occupancy assay, one possible endpoint is the fraction of available target occupied at a specified time. That definition requires the assay to distinguish occupied from unoccupied target in the stated biological context. It does not, by itself, define a clinical response or establish that the target is responsible for a downstream effect.

Response

A response estimand identifies a defined outcome, its direction and units, measurement time, and summary. For instance, it might be the mean change from baseline in a biomarker at 24 hours, or the proportion of patients whose symptom score improves by a prespecified amount by week four. If the goal is a treatment effect rather than a descriptive change, the comparison condition and population must also be defined. A within-group change alone does not isolate an effect of treatment from time trends, disease course, or other influences.

Worked Examples: Make Each Quantity Concrete

What Can Be Transported, and What Needs Human Evidence?

Transportability is not a label attached to a type of variable. It is a claim that a defined relationship or quantity remains applicable across species under specified conditions. A source-species estimate can be measured accurately and still be a poor basis for a human prediction if the relevant physiology, binding, assay, disease state, or exposure-response relationship differs.

Some elements are inherently portable by definition—for example, the unit conversion from milligrams to micrograms—while biological values and relationships generally require evidence and assumptions. A measured animal clearance is an animal estimand. A human clearance prediction is a new target quantity produced by a model or other evidence, with uncertainty. Similarly, an assay-defined engagement value may be transported only if its biological interpretation and measurement are sufficiently comparable; the numeric percentage alone does not establish that comparability.

A useful distinction is between transporting a measurement definition and transporting a biological value or relationship. The former may be feasible after checking analytical comparability. The latter requires a substantive argument about species, population, regimen, and mechanism. Human data are particularly important when the target quantity depends on human-specific factors that the source data cannot identify, or when plausible alternative assumptions lead to materially different predictions.

Justifying a Translation Claim

The reasoning can be organized as a sequence. This is not a substitute for a mechanistic or statistical model; it makes clear what that model must connect.

This sequence also helps distinguish an estimand from an estimate. The estimand is specified before calculation; an estimate is a result obtained from observed data or a fitted model. A model-based human prediction can be reported as an estimate of the human target estimand, but it should not be described as a directly observed human value.

Derivation: Why Mean Unbound Exposure Needs Care

Suppose the target is the population mean unbound AUC over \(0\) to \(T\). For individual \(i\), define \(AUC_{u,i}=\int_0^T f_{u,i}(t)C_{tot,i}(t)\,dt\). The population mean is the mean of these individual integrals. If every individual has the same constant fraction \(f_u\), then linearity of integration gives the familiar simplification. If fractions differ between individuals, the mean of the products generally cannot be replaced by the product of the means.

At each time, \(E[f_u(t)C_{tot}(t)]\) equals \(E[f_u(t)]E[C_{tot}(t)]\) only when the relevant covariance is zero. In general,

Thus, multiplying a mean unbound fraction by a mean total AUC can miss the target population mean unbound AUC when fraction and concentration vary together. The population and summary must be defined, and the calculation must match that definition. This is one reason individual-level joint PK and binding information can matter for a translational estimand.

Diagnostics, Alternatives, and Common Pitfalls

A translation is becoming unreliable when the source and target definitions do not align or when the model’s assumptions are not supported. Check for mismatched plasma versus blood quantities, different assay matrices, incompatible time windows, dissimilar dosing histories, and populations with different disease or covariate profiles. For an exposure estimand, examine whether the sampling and terminal phase support the stated AUC interval. For engagement and response, check whether the assay is valid at the measured time and whether the endpoint has the same meaning across settings. A good numerical fit to source data does not show that the model is informative outside the observed species or conditions.

The main alternatives answer different questions. Directly transporting a value is simple but defensible only when comparability is strongly supported. A mechanistic or empirical bridge model makes the cross-species assumptions explicit, but its prediction remains conditional on the model and the data identifying its relevant components. Collecting human PK, binding, engagement, or response data can address uncertainties the preclinical data cannot resolve, although the needed evidence depends on the target estimand. These approaches can be combined: human data may calibrate a bridge model rather than replace it.

Common mistake: “The animal showed 65% target engagement, so humans will have 65% engagement at the same dose.” This reasoning treats an observed source-species percentage as a transportable human value and ignores differences in exposure, target biology, tissue access, timing, and assay interpretation. The correction is to define the human engagement estimand and identify evidence or assumptions linking the animal measurement to it. A matching dose does not establish matching exposure or engagement.

A related limitation is that a narrowly defined estimand can still be difficult to estimate. Precise wording does not create missing human data, validate an assay, or eliminate structural uncertainty. It does make those gaps visible and prevents a prediction from silently changing its population, endpoint, or time horizon.