Broad-to-Biomarker Selection in a Seamless Trial

Why a Broad-to-Biomarker Transition Needs Precise Claims

A phase 2 study may enroll a broad clinical population while measuring a baseline biomarker that could identify people more likely to benefit. At the transition to phase 3, the design may continue enrollment only in the biomarker-positive subgroup. This can concentrate resources on a population with promising early results, but it changes the population to which the final treatment claim applies.

The essential design task is therefore not simply to choose a subgroup. It is to define the target population, the treatment effect being estimated, and the hypothesis that supports each intended claim. A result in biomarker-positive participants does not automatically establish benefit in the broad population, and it does not by itself establish that the biomarker predicts who benefits more.

These definitions use a causal contrast, but estimation requires an appropriate design. In a randomized trial, treatment assignment must be randomized within the population used for the analysis, and the endpoint, analysis time, treatment contrast, and biomarker rule must be specified. The biomarker should be measured before treatment assignment, using a reliable assay and a fixed positivity threshold. If the endpoint is lower-is-better, define the contrast or direction so that the sign of benefit is unambiguous.

Write the Hypothesis for the Population You Intend to Claim

For a one-sided superiority claim in a higher-is-better outcome, the biomarker-positive efficacy hypotheses are \(H_0:\theta_+\leq 0\) versus \(H_a:\theta_+>0\). A broad-population claim instead concerns \(H_0:\theta_{\mathrm{all}}\leq 0\) versus \(H_a:\theta_{\mathrm{all}}>0\). The hypotheses are not interchangeable: they refer to different populations and generally to different treatment effects.

If the intended claim is that the biomarker is predictive of differential treatment benefit, the relevant contrast is an interaction. On the additive scale for the endpoint, the hypotheses are \(H_0:\theta_+-\theta_-=0\) versus an alternative such as \(H_a:\theta_+-\theta_->0\). The scale and direction must be prespecified. A significant treatment effect in the positive subgroup and a nonsignificant effect in the negative subgroup do not, by themselves, establish a significant interaction.

The broad-population effect can be related to subgroup effects only when the biomarker groups partition the target population and the effects are defined on the same scale and under compatible conditions. For a mean difference, the population-level identity is:

This follows from the law of total expectation. Applying it to each treatment potential outcome and subtracting gives:

The weights must reflect the intended broad target population. If the phase 3 sample is enriched, its observed biomarker prevalence may not represent that target population. Moreover, this identity does not let a positive-subgroup result alone establish the broad effect: the prevalence and the effect in the other subgroup also matter.

Selection at the Boundary and Confirmatory Validity

At the phase boundary, the population choice may depend on phase 2 evidence. That choice is part of the design, not a license to redefine the claim after seeing the results. The biomarker definition, candidate populations, selection rule, and hypotheses should be set in advance. The phase 3 analysis must then account for how the population was selected and whether phase 2 data contribute to the confirmatory test.

If phase 2 results are used only to select a population and the confirmatory evidence comes from independent post-transition data, the testing plan still has to address selection among possible claims. If phase 2 data are also combined with later data, a standard fixed-design analysis that ignores the adaptive selection may not control the type I error rate. The design must use a method valid under the planned selection rule, such as an appropriate combination-test or conditional-error approach, with its error allocation and hypotheses specified before the confirmatory analysis.

This is especially important when the trial could select between a broad-population claim and a biomarker-positive claim. These are distinct hypotheses. A prespecified testing strategy must protect the intended error rate across the claims that may be made; selecting whichever result looks strongest and then applying an unadjusted test does not provide that protection.

Example of a Predictive-Biomarker Claim

A treatment effect within one subgroup and an interaction answer different questions. The following screening calculation illustrates the distinction. It is not a confirmatory analysis: the values are hypothetical phase 2 estimates used to clarify which hypothesis would match each claim.

Specify the Family of Possible Claims

A transition plan should list every population and claim that could be selected, along with the null hypothesis for each. For example, the design might allow a claim of efficacy in the biomarker-positive population, a broad-population efficacy claim, or a differential-benefit claim. It should state which claims are confirmatory, whether more than one may be made, and how the type I error rate is controlled across them.

The following hypothetical allocation illustrates the bookkeeping, not a universal testing recipe. Suppose a design permits two one-sided confirmatory claims at a familywise significance level of \(0.025\): efficacy in the biomarker-positive group and differential benefit by biomarker status. A simple Bonferroni allocation assigns \(0.0125\) to each claim because \(0.0125+0.0125=0.025\). If both claims are tested, this allocation bounds the probability of at least one false rejection by \(0.025\), provided each component test is valid at its assigned level.

In an adaptive seamless trial, “each component test is valid” is a substantive condition. The tests must respect the selection rule and the use of phase 2 information, not merely use the displayed cutoffs. Other prespecified multiplicity procedures may use the error budget differently, but the target claims and their null hypotheses still need to be explicit.

Diagnostics, Failure Modes, and Design Choices

Before the transition, check that biomarker status is available for the participants whose outcomes inform selection, that the assay and threshold are stable, and that the analysis uses the planned population definitions. Compare the biomarker prevalence in the phase 2 population with the intended broad target population. Review subgroup sample sizes and uncertainty: a large point estimate from a small subgroup can be unstable, while an imprecise interaction can make a predictive claim unsupported even when one subgroup appears promising.

A broad-to-biomarker switch is not the only design choice. Continuing broad enrollment preserves direct information across biomarker groups and may support a broad claim, but can spend resources on groups with less apparent benefit. Enrichment focuses enrollment on the selected subgroup and can increase efficiency for a claim about that subgroup, but narrows generalizability and leaves the excluded group less directly studied. Continuing both populations can preserve the ability to assess subgroup and broad effects, at the cost of additional sample size and a more complex multiplicity plan.

Key limitations include assay misclassification, changes in biomarker prevalence across settings, and uncertainty about whether the subgroup effects will transport to the eventual treatment population. A biomarker can be prognostic—associated with outcome regardless of treatment—without being predictive of a different treatment effect. The latter requires evidence on a treatment-by-biomarker contrast, not merely a comparison of outcome levels between biomarker groups.