From a Dose Question to a Model Boundary
A biomarker-guided dose decision is not simply a request to predict an outcome from a biomarker. It asks whether a rule for assigning doses using biomarker information improves the decision compared with a specified alternative. A QSP boundary should therefore begin with the decision itself: the population, available choices, outcome horizon, and consequences that will distinguish a better policy from a worse one.
Consider a hypothetical development question: among adults with a specified disease who are eligible for treatment, should a baseline biomarker test guide a choice between 20 mg and 40 mg of a drug? The analysis will compare the biomarker-guided policy with giving 20 mg to everyone and giving 40 mg to everyone. The numerical examples below are illustrative, not evidence about a real drug or disease.
For this question, define the rule before building the model: use 40 mg if the baseline test is positive and 20 mg if it is negative. The comparator policies are fixed 20 mg and fixed 40 mg. The outcome horizon is 12 weeks. A decision criterion must also state how benefit and harm are weighed; without it, “best dose” is not a complete target.
Specify the Population, Policy, and Outcomes
The target population might be adults meeting the disease definition, eligible for either dose, and receiving the treatment in a specified clinical setting. State relevant restrictions, such as prior treatment or organ-function criteria, if they define who the recommendation is for. The biomarker must be measured at a stated decision time, before the dose is assigned. A recommendation for this population does not automatically apply to children, patients with excluded comorbidities, or a different care setting.
The policy is a mapping from information available at the decision time to an action. Let \(B\) denote the underlying baseline biomarker category, \(T\) the test result, and \(A\) the assigned dose. The intended rule is \(A=40\) mg when \(T\) is positive and \(A=20\) mg when it is negative. In practice, \(T\) may not identify \(B\) perfectly, so the assay is part of the decision pathway rather than an interchangeable label for the biological state.
For a transparent comparison, suppose benefit is the probability of clinical response by week 12, and harm is the probability of a specified adverse event by week 12. Define a simple decision score \(U=P(\text{response})-2P(\text{adverse event})\). This score assigns an adverse event twice the weight of a response event; that is a decision assumption, not a biological law. A different stakeholder valuation could lead to a different policy.
Worked Example: Writing the Target Comparison
Suppose the decision team wants an initial model boundary for the hypothetical setting above. State the target comparison precisely enough that two analysts would evaluate the same policies.
Population: adults eligible for either dose in the specified treatment setting, with baseline biomarker testing before treatment assignment.
Policies: (1) assign 40 mg for a positive test and 20 mg for a negative test; (2) assign 20 mg to all; (3) assign 40 mg to all.
Outcomes and horizon: response and the specified adverse event, each assessed by week 12. The decision score is \(U=P(\text{response})-2P(\text{adverse event})\).
Target estimand: the difference in expected decision score under the biomarker-guided policy versus each fixed-dose policy in this target population. This is a policy comparison, not merely a comparison of outcomes among people observed to have different biomarker values.
Trace the Causal Pathways and Set the Interfaces
A useful mechanistic boundary follows the pathways that could make the policies differ. A candidate causal chain is: assigned dose affects systemic exposure; exposure affects target engagement; target engagement changes disease-relevant pathway activity; pathway activity contributes to clinical response. In parallel, exposure and pathway effects may contribute to adverse events. Baseline biomarker status may alter target abundance, pathway state, or the relationship between target engagement and response. These are hypotheses to assess, not facts established by drawing arrows.
The model should distinguish a biological biomarker state from its measured test result when test error could change assignment. It should also distinguish dose assignment from exposure: the dose is a decision, whereas exposure is a downstream quantity affected by pharmacokinetics and possibly patient characteristics. The exposure interface can be supplied by a linked PK model or represented by a justified exposure input. Do not represent the same exposure process in two linked components.
A practical boundary can include baseline biomarker status, drug exposure, target engagement, a small number of relevant pathway states, and the outcome processes needed to predict response and harm over 12 weeks. The exact biological states depend on available evidence and decision relevance. Age, organ function, concomitant treatment, or disease severity might enter as population descriptors, model inputs, or effect modifiers if they influence the policy comparison. Their role should be explicit.
Processes that do not plausibly affect the comparison over the chosen horizon can be excluded, but their exclusion is a testable assumption. For example, a long-term disease progression process may be outside a 12-week endpoint boundary if it cannot materially affect response, harm, or dose ranking within that period. Conversely, excluding a feedback loop is not justified merely because it is difficult to model: if it changes exposure, pathway response, or toxicity differently across the two doses, it may change the decision.
Worked Example: Translating a Causal Pathway Into an Interface
Suppose the team has a validated exposure model but is constructing a QSP model for the biomarker-to-response pathway. Decide what crosses the component boundary.
The dose assignment is passed to the exposure model, which returns an exposure-time profile for each simulated individual. That profile is an input to the QSP model. The QSP model represents target engagement and downstream pathway states, then returns predicted response and adverse-event probabilities by week 12.
This interface is coherent if the exposure model accounts for the relevant dose-to-exposure processes and the QSP model does not separately calculate the same processes. If renal function affects exposure, it must be represented in the exposure component or passed through an interface that already incorporates it; silently omitting it from both components would leave a causal path unmodelled.
The biomarker test result is a second interface: the policy uses the test result to choose a dose, while the biological biomarker state may modify the QSP response. If the test is imperfect, a measurement model links the biological state to the result. Treating the test result as the biological state is defensible only when misclassification is negligible for the decision or has been shown not to alter the recommendation.
Calculate a Policy Comparison
The policy comparison can be written as an average across biomarker groups. Let \(p_b\) be the prevalence of group \(b\), and let \(U(a,b)\) be the expected decision score if dose \(a\) is given to group \(b\). For a policy \(g\), which assigns a dose \(g(b)\), its expected score is the population-weighted average of the group-specific scores. This calculation assumes the group proportions and conditional outcome predictions apply to the stated target population, and that the score adequately represents the decision objective.
A useful justification follows by expanding the definition. In each biomarker group, the policy assigns one dose and incurs that group’s expected score. The groups partition the target population, so the law of total expectation gives the overall expected score as the sum of each group score weighted by its population proportion. Comparing policies therefore requires both the dose-specific predictions and the distribution of biomarker groups.
Worked Example: Comparing the Three Policies
Assume, hypothetically, that 40% of the target population is biomarker-high and 60% is biomarker-low. The model predicts the following week-12 response and adverse-event probabilities. Calculate each score using \(U=P(\text{response})-2P(\text{adverse event})\).
| Biomarker group | Dose | Response probability | Adverse-event probability | Score \(U\) |
|---|---|---|---|---|
| High | 20 mg | 0.55 | 0.08 | 0.55 − 2(0.08) = 0.39 |
| High | 40 mg | 0.78 | 0.18 | 0.78 − 2(0.18) = 0.42 |
| Low | 20 mg | 0.50 | 0.07 | 0.50 − 2(0.07) = 0.36 |
| Low | 40 mg | 0.52 | 0.15 | 0.52 − 2(0.15) = 0.22 |
The guided policy gives 40 mg to the high group and 20 mg to the low group: \(U(g)=0.40(0.42)+0.60(0.36)=0.168+0.216=0.384\).
For fixed 20 mg, \(U=0.40(0.39)+0.60(0.36)=0.156+0.216=0.372\). For fixed 40 mg, \(U=0.40(0.42)+0.60(0.22)=0.168+0.132=0.300\). Under these hypothetical predictions and this score, the guided policy exceeds fixed 20 mg by 0.012 score units and fixed 40 mg by 0.084 score units. These are differences in the defined decision score, not percentage-point increases in response.
The guided policy’s advantage over fixed 20 mg is modest. The result alone does not establish that the policy should be used: it depends on the population proportions, predicted outcomes, test performance, and the chosen valuation of benefit and harm.
Check the Boundary Before Trusting the Ranking
Boundary checks should target assumptions capable of changing the policy ranking, not only whether the model runs or fits a calibration dataset. Vary uncertain mechanisms, population characteristics, interface assumptions, and excluded processes over plausible ranges. Recalculate the policy comparison each time. If a plausible change reverses the preferred policy, the current boundary is not yet robust enough to support an unconditional recommendation.
Test performance is one concrete example. Suppose the assay has 90% sensitivity and 90% specificity for the high biomarker state, with high-state prevalence 0.40. Then the probability of a high-state person testing positive is \(0.40(0.90)=0.36\); the probability of a low-state person testing positive is \(0.60(0.10)=0.06\). If the test-positive rule assigns 40 mg and test-negative assigns 20 mg, its score difference from fixed 20 mg is:
Under these assumptions, the test-guided policy still has a higher expected score than fixed 20 mg, but the advantage falls from 0.012 to 0.0024 score units. This calculation assumes the stated sensitivity and specificity apply in the target population and that outcomes conditional on true biomarker group and dose are as listed. It also illustrates why a model should not equate biomarker classification with a perfect assay without justification.
Other checks include comparing predictions with observed dose-response patterns, checking whether the model reproduces relevant biomarker-stratified outcomes, and testing whether plausible omitted feedback or delayed effects alter the ranking. Inspect whether predictions behave sensibly at the boundaries of the simulated population and dose range. A boundary may be inadequate if the model requires implausible parameter changes to match data, predicts effects inconsistent with known biology, or changes its recommendation under small, plausible perturbations.
Alternatives, Limitations, and a Common Pitfall
A direct empirical comparison of randomized dose policies can estimate outcomes under the tested conditions with fewer mechanistic assumptions, but it may not cover every biomarker subgroup, dose, or population of interest. A statistical biomarker-by-dose model can estimate subgroup differences efficiently but may not represent the underlying pathways or support mechanistic extrapolation. A QSP model can connect dose, exposure, biomarker biology, and outcomes, which may help explore untested conditions; that advantage depends on the credibility of its mechanisms, interfaces, and validation evidence. These approaches can complement one another rather than serve as mutually exclusive choices.
The boundary limits what the model can answer. A 12-week model does not establish long-term benefit or harm if those outcomes are outside its states and evidence. A recommendation in one population does not establish transportability to a different population. Uncertainty in model structure, parameter values, test performance, and decision-maker preferences should be reported; a single predicted ranking can conceal substantial uncertainty.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Why must the biomarker-guided policy be compared with explicitly stated alternatives?
- With biomarker-high prevalence 0.30, suppose scores are 0.41 for high-group 40 mg, 0.38 for high-group 20 mg, and 0.35 for low-group 20 mg. What is the guided-policy score if low biomarker status receives 20 mg?
- In the preceding question, what is the fixed-20-mg score if the high-group 20-mg score is 0.38?
- Name one reason to model the measured test result separately from the biological biomarker state.
- What does it imply if a plausible change to an excluded feedback process reverses the preferred policy?
Solutions. (1) A policy is only preferable relative to a comparator and a stated decision criterion; another comparator or valuation of harm may change the ranking. (2) The score is \(0.30(0.41)+0.70(0.35)=0.123+0.245=0.368\). (3) The fixed-dose score is \(0.30(0.38)+0.70(0.35)=0.114+0.245=0.359\), so the guided policy exceeds it by 0.009 score units under these assumptions. (4) The test may misclassify the biological state, and that error can change which dose the policy assigns. (5) The boundary may omit a decision-relevant mechanism; its effect should be represented, bounded, or included in uncertainty analysis before relying on the recommendation.