Setting Quantitative Go Or No-Go Criteria for Early Trials

From Mechanistic Predictions to a Decision Rule

An early trial is useful when its outcomes can change a development decision. QSP predictions can make that decision quantitative by linking a dose to target engagement, target engagement to a biomarker, and the biomarker to a clinical endpoint. The decision rule should state in advance what level of evidence across this chain is sufficient to proceed. A mechanistic signal alone may not support a clinical benefit; a clinical estimate alone may not establish that the intended mechanism was engaged.

The purpose here is to derive and apply operational go/no-go criteria, not to treat a model prediction as a guarantee about patients. The numbers below are hypothetical. They illustrate how a team might connect a model’s predictions to measurable trial criteria; actual thresholds require justification for the disease, assay, endpoint, dose, and development decision.

For the worked framework, let \(E\) denote fractional target engagement, bounded by \(0\le E\le1\). Let \(B(E)\) denote predicted biomarker reduction in percentage points, and let \(C(B)\) denote predicted clinical improvement in endpoint units, where larger positive values represent greater benefit. These definitions fix the direction of every comparison: engagement, biomarker reduction, and improvement are all treated as positive quantities.

Deriving a Mechanistically Linked Threshold

Suppose a hypothetical QSP model predicts the following relationships at the chosen assessment time:

Here, 30 percentage points is the asymptotic maximum of the biomarker curve as \(E\) increases without bound; it is not the maximum over the allowed engagement range. Because \(B(E)\) increases with \(E\), its maximum for \(0\le E\le1\) is \(B(1)=30/(1.20)=25\) percentage points. Likewise, the clinical-response curve has an asymptotic maximum of 1.8 endpoint units. These bounds matter: a required response beyond the model’s attainable range cannot be met by choosing a higher dose within the model.

Assume the development team defines a clinically meaningful model-based goal of at least 1.0 endpoint unit of improvement. The linked threshold can be derived backward, from clinical response to biomarker change and then to engagement. That is more informative than selecting an engagement cutoff without checking whether it corresponds to the intended clinical goal.

Since \(B(E)\) is increasing on the allowed range, solving \(B(E)\ge12.5\) gives the minimum engagement:

Thus, the model predicts that engagement of at least 0.1429, or 14.29% on a percentage scale, is required to reach the 1.0-unit clinical goal. At this boundary, \(B=12.5\) percentage points and \(C=1.0\) endpoint unit. The threshold is feasible because it lies in the permitted engagement range. At \(E=0.15\), for example, \(B=30(0.15)/(0.20+0.15)=12.86\) percentage points and \(C=1.8(12.86)/(10+12.86)=1.013\) endpoint units, just above the goal.

Worked Example: Check Whether the Clinical Goal Is Attainable

From Individual Predictions to a Population Criterion

A threshold for one simulated individual does not say how many people in the intended trial population are predicted to reach it. One population-level criterion is the fraction of virtual individuals whose engagement at the planned dose meets or exceeds \(E=0.1429\). This fraction is interpretable only if the virtual population represents the intended trial population and any unequal population weights are handled appropriately. It is a model-based prediction, not an observed clinical-trial proportion.

Prespecifying Trial Gates and Applying Them

A trial decision rule needs measurable criteria, not just a model-derived target. One illustrative conjunctive rule could require all three of the following at a prespecified dose and assessment time: at least 70% of participants meet the target-engagement threshold of \(E\ge0.1429\); the placebo-adjusted mean biomarker reduction is at least 12.5 percentage points; and the placebo-adjusted mean clinical improvement is at least 1.0 endpoint unit. “Placebo-adjusted” means the estimated improvement in the active-treatment group relative to the comparator, with the sign defined so that positive values favor treatment.

The engagement fraction checks whether enough participants reach the model-derived mechanistic threshold. The biomarker criterion checks whether the expected downstream pharmacodynamic change is seen at the group level. The clinical criterion checks whether the trial’s endpoint estimate reaches the chosen development goal. The biomarker and clinical cutoffs are decision thresholds, not claims that the corresponding sample estimates are exact or free of sampling error.

Assumptions, Diagnostics, and Alternatives

The numerical rule depends on assumptions that should be checked before it is used. The model’s exposure and engagement predictions must correspond to the proposed dose and schedule; the modeled biomarker and clinical endpoint must match the assay, endpoint definition, and assessment time used in the trial; and the population used to calculate the virtual fraction must be relevant to the intended trial population. The response relationships must also be appropriate over the range used to derive the thresholds. If the model predicts a benefit at a time point different from the trial’s primary assessment, the numerical mapping is not directly comparable.

Diagnostics include checking units and directions, plotting the predicted relationships over the full allowed range, and verifying that every calculated threshold lies in that range. In this example, \(B(1)=25\), not 30 percentage points; therefore a biomarker requirement above 25 percentage points cannot be met within the specified engagement domain. A target clinical benefit above the response curve’s 1.8-unit asymptote is also unattainable under the assumed relationship. A negative value, an abrupt discontinuity, or a threshold that changes materially under small numerical rounding should prompt a review of the model, calculation, and definitions before a decision is made.

A conjunctive rule is not the only possible design. A single clinical-endpoint gate is simpler, but does not establish whether the intended mechanism was engaged. A single mechanistic gate may support biological activity while leaving clinical relevance unresolved. A weighted composite can summarize several outcomes, but its weights can allow a strong result on one measure to offset a failure on another; that may be undesirable when each link in the causal chain is essential. The conjunctive rule makes each requirement explicit, at the cost of being less forgiving when one measure narrowly misses its cutoff.

A cutoff applied to the mean engagement value is another alternative. However, a mean can conceal a mixture of participants with high and low engagement. The fraction above a mechanistically derived threshold directly expresses the proposed target-attainment criterion, although it still depends on accurate measurement and a representative analysis population. No rule removes sampling variability or limitations in the model-to-trial mapping. The numerical cutoffs should therefore be justified and prespecified, and their precision and operating characteristics considered in trial planning.

A common mistake is to treat the model’s asymptotic Emax as the maximum achievable output in the allowed domain. Here, calling 30 percentage points the maximum predicted biomarker reduction over \(0\le E\le1\) is incorrect: it is the limit as engagement grows without bound, whereas the bounded-domain maximum is 25 percentage points. The correction is to state the domain, evaluate the function at its boundary, and use that value when judging feasibility. A second pitfall is to let a positive biomarker result override a failed clinical gate after seeing the data. The correction is to apply the prespecified rule consistently and report which criterion failed.