Start With the Decision, Not the Model
A development team may ask whether to increase a dose, change a dosing schedule, or predict an outcome in a population not yet studied. Each question may involve drug exposure and response, but that does not mean it requires the same model. The useful starting point is to state the action that the analysis will inform, the conditions under which it must predict, and what could happen if the prediction is wrong.
An empirical PK/PD model relates drug concentration or another exposure measure to a pharmacodynamic response using a specified mathematical relationship. It may be well suited to predicting responses within the conditions represented in its data. A QSP model represents biological mechanisms and interacting states that connect drug action to downstream responses. That representation can support some predictions beyond the conditions directly observed, but only when the relevant mechanisms and their links are supported by evidence. The distinction is not “simple versus good” or “complex versus accurate”: it is a difference in what each model represents and what its predictions rely on.
The earlier material on causal mechanisms versus empirical exposure–response fits and on evidence chains from drug action to clinical outcome provides useful distinctions: a fitted relationship does not, by itself, establish a causal pathway, and a mechanistic diagram does not establish that every link is supported. Here those distinctions become a practical model-selection question.
A Decision-Based Selection Framework
Before choosing an approach, specify four things. First, name the decision and the model output that would inform it. Second, identify the intended prediction conditions: population, dose, schedule, time horizon, and endpoint. Third, describe how far those conditions lie from the supporting data. Fourth, list the evidence each candidate model would need. This is not a numerical score. It is a way to expose whether a model’s assumptions fit the use being proposed.
State the available actions and the outcome that distinguishes them. A model is relevant only if its output can inform that choice.
Interpolation within observed exposure and treatment conditions usually places fewer demands on a model than changing a schedule, population, species, or biological context.
For a QSP prediction, identify the measured support for each mechanism and link needed to reach the decision-relevant outcome. Do not treat missing links as established merely because they appear in equations.
A consequential decision calls for stronger support, more explicit uncertainty, and possibly additional data. High stakes alone do not make a mechanistic model reliable.
A useful justification for this framework comes from the decision itself. Suppose \(a\) is an available action, \(m\) is a plausible model, and \(U(a,m)\) is the utility of taking action \(a\) if model \(m\) describes the relevant outcomes. If every well-supported candidate model favors the same action, choosing a more elaborate model may not change the decision. If candidate models favor different actions, the disagreement matters: the team needs to examine the assumptions and evidence driving it before relying on one prediction.
This is a decision principle, not a guarantee that any model is correct. When plausible models yield different \(a_m\), a model-selection discussion should identify the uncertain assumptions responsible for the difference. If the evidence cannot resolve them, a new experiment or a narrower decision may be more defensible than choosing the model with more biological detail.
Worked Example: Interpolating Within a Studied Exposure Range
Hypothetical setting. A team is choosing between two doses for the same patient population, treatment schedule, and assessment time already represented in its exposure–response data. The decision concerns whether the mean disease-marker value is expected to fall below 85 units. An empirical model fitted to those data is \(E(C)=100-40C/(2+C)\), where \(C\) is concentration in mg/L and \(E(C)\) is the expected marker value in units. The observed exposure range spans 1 to 4 mg/L.
For the candidate exposure \(C=1.5\) mg/L, the prediction is an interpolation within the observed range. Substituting the concentration into the fitted relationship gives:
Under the model, the expected marker value at this exposure is about 82.86 units, below the decision threshold of 85 units. For context, the same curve gives \(E(1)=100-40/3\approx86.67\) units and \(E(4)=100-160/6\approx73.33\) units, consistent with evaluating \(1.5\) mg/L between exposures represented in the data.
An empirical PK/PD model is a reasonable candidate here if the data support the fitted relationship and the population, schedule, endpoint, and assessment time are sufficiently aligned with the proposed use. The calculation does not show that changing exposure caused a change in the marker, nor does it establish predictions for a different schedule or population. Those would be different uses requiring further justification.
When the Question Requires Mechanistic Extrapolation
A QSP model becomes more relevant when the decision depends on biological features that an exposure–response curve does not represent adequately. Examples include a change in schedule that alters pathway recovery, a perturbation of a biological component, or translation across experimental systems where the underlying mechanisms matter. The key requirement is not simply that such features sound biological. Evidence must support the states and links needed to predict the endpoint under the new conditions.
Even a QSP model does not automatically solve an extrapolation problem. Its prediction depends on the relevance of its mechanisms, parameter values, and connections from drug action to outcome. If a decision relies on an unsupported link—for example, assuming that a biomarker change reliably predicts a clinical effect—the mechanistic label does not repair the gap. The earlier material on recognizing unsupported links in mechanistic extrapolation is especially relevant in such cases.
Worked Example: A Schedule Change With Pathway Persistence
Hypothetical setting. A team has studied a continuous infusion that maintains concentration at 1 mg/L for 24 hours. It is considering a 12-hour infusion at 2 mg/L followed by 12 hours without drug. Both schedules have the same 24-hour area under the concentration–time curve: for continuous infusion, \(1\ \text{mg/L}\times24\ \text{h}=24\ \text{mg}\cdot\text{h/L}\); for the proposed schedule, \(2\ \text{mg/L}\times12\ \text{h}+0\times12\ \text{h}=24\ \text{mg}\cdot\text{h/L}\).
Equal area under the curve does not imply equal pathway drive when the concentration–effect relationship is nonlinear. For illustration, suppose the supported proximal relationship is \(I(C)=C/(1+C)\), with \(C\) in mg/L and \(I(C)\) a dimensionless inhibition measure. Under continuous infusion, the 24-hour mean drive is \(I(1)=1/2\). Under the proposed schedule, the mean is:
These values are not clinical endpoint predictions. They show why equal exposure area alone need not imply equal pathway input. The team also has hypothetical assay evidence that the downstream pathway state recovers slowly after drug withdrawal. A candidate QSP representation could therefore include a pathway state \(S(t)\), with input from \(I(C(t))\) and loss at a rate determined by a recovery parameter. Whether that model is suitable depends on evidence supporting the state, its recovery behavior, and its relationship to the clinical endpoint.
If the decision depends on the consequence of a drug-free interval, a model using only total exposure may omit a decision-relevant feature. A QSP model is a plausible candidate because the prediction requires representing time-varying pathway behavior. But if recovery has not been measured, the model may not support a reliable schedule recommendation; targeted pathway time-course data may be needed before relying on its extrapolation. A time-varying empirical PK/PD model could also be adequate if it has relevant schedule data. The choice is determined by the evidence and intended prediction, not the model label.
When Neither Model Supports the Requested Prediction
Sometimes the required extrapolation extends beyond both the empirical data and the mechanisms that can be supported. In that situation, the disciplined choice is not necessarily to force a choice between PK/PD and QSP. An empirical model may still summarize observed relationships, while a QSP model may organize hypotheses, but neither should be presented as establishing a prediction for the unsupported context.
Worked Example: A High-Consequence Prediction With a Missing Link
Hypothetical setting. In an early study, three exposure bands have mean biomarker changes of 5, 12, and 18 units, and corresponding mean clinical endpoint improvements of 1, 2, and 3 points. The team wants to use these observations to predict clinical benefit at 4 mg/L in a new population, where exposures above 2 mg/L and the biomarker-to-endpoint relationship have not been studied. A wrong prediction could lead to a consequential treatment recommendation.
The observed summaries do not establish that the biomarker change causes or reliably predicts clinical improvement. An empirical exposure–response fit could describe the measured association across the observed bands, but extrapolating it to 4 mg/L and a new population would require assumptions not tested by those observations. A QSP model could represent a proposed pathway from exposure through the biomarker to the clinical endpoint, but its prediction would depend on the unverified biomarker-to-endpoint link and on how that link transfers to the new population.
The appropriate conclusion is that neither candidate currently justifies the requested clinical prediction. The team could use the empirical model to summarize the observed range, clearly limit its interpretation, and seek data that test the missing link and the new exposure range. The high consequence raises the standard of evidence; it does not transform an unsupported mechanistic assumption into a reliable QSP prediction.
Compare the Approaches and Diagnose a Poor Fit
| Question | Empirical PK/PD is a stronger candidate when… | QSP is a stronger candidate when… |
|---|---|---|
| What is being represented? | A fitted relationship between exposure and response is sufficient for the decision. | Biological states and causal links are needed to represent the decision-relevant pathway. |
| Where is the prediction? | It is close to the observed population, exposure range, schedule, and endpoint. | It involves a changed context for which supported mechanisms can inform extrapolation. |
| What evidence is needed? | Relevant exposure and response data that support the fitted relationship. | Evidence for the mechanisms, parameters, and links used to reach the predicted outcome. |
| What if support is inadequate? | Restrict the use to the supported range and conditions. | Restrict the use, test uncertain links, or collect additional mechanistic evidence. |
Diagnostics should be tied to intended use. For an empirical model, inspect whether the proposed exposure and context are represented in the data and whether its predictions describe relevant observations. For QSP, trace each decision-relevant prediction back through the states and links that produce it; identify which links are measured, inferred, or simply assumed. For either approach, compare predictions across plausible assumptions. If a modest change in a weakly supported assumption reverses the recommended action, the decision is fragile even if the model fits available data.
A common mistake is: “The decision is important, so we should use QSP because it is more mechanistic.” Importance does not validate mechanisms. The correction is to ask whether the decision requires a mechanistic extrapolation and whether evidence supports the mechanisms that extrapolation uses. If not, collect relevant evidence, narrow the model’s scope, or defer the unsupported prediction.
A second limitation is that these approaches are not mutually exclusive. An empirical PK/PD relationship may be one component of a QSP model, and a QSP model may be used to organize data without supporting every requested prediction. A model can also be useful for one decision and unsuitable for another. State what the model is being used to predict and avoid extending conclusions beyond that context.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A team predicts response at an exposure between two well-observed exposures, in the same population and schedule. What is the initial model-selection argument, and what must still be checked?
- Two schedules have equal total exposure, but the decision depends on pathway recovery during drug-free periods. What feature might a QSP model represent, and what evidence is needed?
- Why does a high-consequence decision not automatically imply that QSP is the right approach?
- An empirical model fits the observed range, but the requested prediction is in a new population and depends on an unvalidated biomarker-to-outcome link. What is a defensible next step?
Solutions. (1) An empirical PK/PD model is a reasonable initial candidate because the prediction is interpolation under represented conditions. Check that the exposure–response relationship is supported and that the endpoint, time, and population match the intended use. (2) A QSP model might represent a persistent pathway state that recovers over time. Evidence should support the pathway state, its recovery, and the link from that state to the relevant outcome; equal total exposure alone is not enough. (3) High consequences raise the standard for evidence and uncertainty assessment, but do not establish the validity of a model’s mechanisms. (4) Limit the empirical model to its supported use, test the missing link and population transfer, or collect additional data. Do not present an unsupported extrapolation as validated merely by fitting a more mechanistic model.