Biomarker Enrichment in an Adaptive Trial

Why Biomarker Enrichment Requires a Decision Rule

An adaptive trial may begin by enrolling patients across biomarker groups, then use interim outcomes to decide whether later enrollment should continue in the full population or focus on a subgroup. The decision is consequential: enrichment can concentrate enrollment where treatment appears more useful, but it can also exclude patients based on noisy interim evidence. A defensible decision therefore requires a defined treatment-benefit measure, a model for the interim data, and a rule specified before examining those data.

The example here concerns a binary response measured over a fixed follow-up period. The same decision logic can be adapted to other outcomes, but the likelihood and estimand must match the outcome and its follow-up structure. In particular, a response probability at a fixed time is not interchangeable with a time-to-event quantity.

Model and Assumptions

For each arm and subgroup, suppose the interim data consist of \(x_{a,s}\) responders among \(n_{a,s}\) patients with an evaluable response outcome. Under a binomial model, \(X_{a,s}\mid p_{a,s}\sim\operatorname{Binomial}(n_{a,s},p_{a,s})\). A convenient prior is \(p_{a,s}\sim\operatorname{Beta}(\alpha_{a,s},\beta_{a,s})\). The binomial likelihood and Beta prior give the posterior

This model assumes that, within each arm and subgroup, patients have a common response probability and their response outcomes are conditionally independent given that probability. It also assumes that the endpoint definition and assessment time are consistent, that the interim data are correctly classified, and that missing outcomes are handled in a way compatible with the model. Randomization supports treatment comparisons within biomarker strata, but it does not guarantee that these other assumptions hold.

For the calculations below, use independent \(\operatorname{Beta}(1,1)\) priors for the four arm-by-subgroup response probabilities. This prior is uniform on each response probability. Independence is a modeling choice, not a consequence of randomization: it means the model does not borrow information across arms or subgroups. If borrowing is desired, it requires a different model and careful assessment of how much information may move across groups.

Posterior Probability of Clinically Meaningful Benefit

A decision rule often asks whether benefit exceeds a clinically meaningful margin \(\delta\), rather than merely whether benefit is positive. The subgroup-specific quantity is \(P(\Delta_s>\delta\mid\text{data})\). With independent Beta posteriors for treatment and control, this probability can be calculated by numerical integration. If \(f_T\) is the treatment posterior density and \(F_C\) is the control posterior cumulative distribution function, then for \(\delta\geq0\),

The integral runs over the full support of the treatment response probability; the control CDF is zero when its argument is below 0. This integral accounts for uncertainty in both response probabilities. When posterior distributions are reasonably concentrated and away from the boundaries 0 and 1, a normal approximation to their difference can be a useful transparent calculation. If the posterior means and variances are \(m_T,v_T\) and \(m_C,v_C\), independence gives approximate difference mean \(m_T-m_C\) and variance \(v_T+v_C\). Thus,

where \(\Phi\) is the standard normal cumulative distribution function. This approximation is not the exact Beta-difference probability. For a consequential trial decision, numerical integration or posterior simulation can check its accuracy, especially when subgroup counts are small or posterior mass lies near 0 or 1.

Applying an Adaptive Enrichment Rule

Suppose the protocol specifies that a subgroup qualifies for continued enrollment if its posterior probability of exceeding the 10-percentage-point margin is greater than 0.90. It also specifies that, if only one subgroup qualifies, later enrollment is restricted to that subgroup. The rule is applied to the results just calculated.

A subgroup-specific benefit probability is not, by itself, proof that the biomarker is predictive of differential treatment effect. A biomarker may be associated with response in both arms, making it prognostic, while treatment effects remain similar across subgroups. To support a claim of differential benefit, the analysis should also examine the contrast \(\Delta_+-\Delta_-\), with uncertainty, under a prespecified model. The enrichment rule above is an operational enrollment decision based on separate subgroup criteria; its conclusion should not be overstated as a definitive interaction finding.

Diagnostics, Alternatives, and Limitations

Before acting on interim subgroup results, check that subgroup membership was determined using a prespecified assay and cutpoint, that randomization and outcome assessment were maintained within relevant strata, and that response status is sufficiently complete at the planned assessment time. Review the actual counts and denominators by arm and subgroup. A large imbalance in evaluability, unexpected missingness, or inconsistent response assessment can make a simple binomial analysis misleading. Also assess whether the common-probability assumption is plausible across sites, calendar periods, or clinically important patient groups.

Compare the primary result with reasonable prior choices and, where appropriate, with a model that allows treatment effects to differ across subgroups. Independent priors avoid cross-subgroup borrowing but may leave estimates unstable when data are sparse. Hierarchical models can stabilize subgroup estimates by partial pooling, but may also pull a genuinely different subgroup toward the others. The choice should reflect the trial’s scientific aim and be evaluated before relying on an interim decision.

The main alternatives answer different questions. Continuing all-comer enrollment avoids making a subgroup restriction based on interim evidence, but may spend enrollment on patients with lower expected benefit. A fixed enrichment design selects a subgroup before the trial and does not adapt using interim outcomes. A prespecified adaptive rule uses interim information, but must account for the fact that the rule itself affects who is enrolled and what evidence will accrue. A simple comparison of observed response proportions is not a substitute for posterior uncertainty, and selecting whichever subgroup has the largest observed rate can be particularly unstable.

A common mistake is to reason, “The positive subgroup crossed the threshold and the negative subgroup did not, so the biomarker has been proven to identify responders.” The correction is that the threshold supports only the specific decision stated in the rule, under the model and data available. It does not establish certainty, prove a treatment-by-biomarker interaction, or remove uncertainty about future patients. Enrichment also changes the population represented by later trial data, so conclusions about an all-comer population may require separate justification.