Biomarker use in subgroup analysis versus enrichment trial design as a predictor of oncology trial success.

Abstract

e23122 Background: Biomarkers are widely used in oncology clinical trials for patient stratification and endpoint definition, yet it remains unclear whether trials should i) use an “enrichment design” that treats only biomarker-positive patients or ii) use other designs that treat biomarker-negative patients and use biomarkers for subgroup analysis. We performed the first systematic analysis of biomarker placement across all interventional oncology clinical trials to assess whether enrichment designs are associated with higher trial success rates. Methods: We analyzed 105,203 interventional oncology trials from Trialtrove, a curated clinical trial database. After excluding trials with incomplete biomarker data or indeterminate outcomes, 22,277 trials with 3,711 unique biomarkers remained. For 161 frequently studied biomarkers, such as HER2, CYP3A4, EGFR, biomarker mentions were identified across eight trial documentation fields: Trial Title, Trial Objective, Patient Population, Inclusion Criteria (IC), Exclusion Criteria (EC), Primary Endpoints, Secondary Endpoints, and Trial Results. Biomarker use in IC or EC defined enrichment designs. Each biomarker mention was paired with its surrounding text to determine its semantic role. A subset of biomarker-context pairs was manually labeled as true, incidental, or false mentions, and used to train a BERT-based classifier (94.1% accuracy) to label remaining pairs at scale. For success analyses, user-defined field sets were constructed by grouping one or more documentation fields containing biomarker mentions, while non-field sets had all remaining documentation fields not in the field set. Trial success was defined as positive or early positive outcomes, and success rates were compared between field and non-field sets using chi-square tests with Bonferroni correction. Results: Comparisons between user-chosen field and not-chosen field sets (all other fields) revealed consistent differences in trial success rates. When biomarkers were used in field sets corresponding to patient selection—most notably Inclusion Criteria and Exclusion Criteria—trials consistently showed lower success rates than when the same biomarkers appeared in the other six fields. In contrast, field sets that included downstream trial components suggestive of post-hoc analysis, particularly Trial Results and Patient Population, were associated with significantly higher success rates (e.g. 0.692 vs 0.360 for Trial Results and all other fields, respectively). Overall, trials in which biomarkers were evaluated outside enrollment-focused fields demonstrated greater likelihood of success than trials using biomarkers for patient selection. Conclusions: Future trials may benefit from minimizing biomarker-based patient selection in favor of designs that have biomarker-positive and biomarker-negative subgroups that are compared after treatment.