Enrichment and biomarker-stratified designs

A biomarker can be prognostic, telling you how the disease will behave regardless of treatment, or predictive, telling you whether a particular treatment will work. Only a randomised comparison within marker-positive and marker-negative patients can distinguish the two, and the design chosen decides what will be learned. An enrichment design enrols only marker-positive patients: it is smaller and more likely to succeed, and it produces a drug and a companion test together, but it cannot show that marker-negative patients would not also benefit. A biomarker-stratified design enrols all-comers, randomises within each marker group and pre-specifies a test of whether the treatment effect differs between them. An adaptive enrichment design starts with all-comers and narrows to the marker-positive group at an interim if that is where the effect is.

MAGNITUDE is the corpus example of parallel cohorts: it tested niraparib with abiraterone in first-line metastatic castration-resistant prostate cancer in separate cohorts with and without homologous recombination repair alterations, stopped the marker-negative cohort for futility and met its endpoint in the BRCA cohort, settling a debate that an enriched trial alone could not have settled. TAILORx, MINDACT and RxPONDER are stratification by genomic risk score used to decide who can skip chemotherapy: RxPONDER found that chemotherapy added nothing for postmenopausal women with a low score and one to three positive nodes but improved invasive disease-free survival in premenopausal women, a difference no enriched design would have revealed. IMvigor011 is enrichment by circulating tumour DNA, randomising only the patients whose test was positive.

The failure modes are specific. Enriching on an unvalidated marker can exclude the patients who would have benefited most, and the marker cut-off chosen for the trial becomes the cut-off written into the label. Testing many candidate markers after the fact is subgroup fishing under another name. And a trial powered for the overall population but analysed by marker status is usually underpowered within each stratum, so a marker that looks predictive in a forest plot needs its interaction test, not just two separate p-values.