Adaptive Trial Designs for Assessing Predictiveness of a Biomarker in Early Phase Drug Development

Biomarker identification plays a crucial role in precision medicine, where treatments are tailored to patients' intrinsic factors. Predictive biomarkers can be used to identify patients who are more likely to benefit from a treatment. However, identifying and validating such biomarkers pose substantial statistical and practical challenges, especially in the small sample setting. Several methods have been proposed for assessing whether a biomarker has a predictive effect, including the Average Kolmogorov-Smirnov Approach (AKSA), which has been shown to have improved operating characteristics for inference compared to alternative approaches. Yet power remains limited when sample sizes are small and unbalanced across treatment arms. Power can be improved through adaptive enrichment of the sample size, especially when treatment arms are small and unbalanced. We propose an adaptive design to improve power for detecting a predictive effect when sample sizes are small. Existing adaptive biomarker designs largely focus on enrichment of the trial, for example, adjusting enrolment criteria based on an assessment of the treatment-biomarker interaction at interim, or refining the biomarker threshold used to segregate the population into biomarker positive or negative (i.e., those patients more or less likely to respond to the treatment). Few designs explicitly address the task of formally declaring a predictive effect of a biomarker using an interim analysis. The interim analysis framework enables a more reliable assessment of a biomarker's predictive effect by introducing a pre-planned evaluation at the originally intended sample size of a study which may otherwise be underpowered due to the limited number of patients. Based on the interim results, the trial may (i) already have sufficient evidence to declare a biomarker as predictive, (ii) be terminated early due to a clear lack of evidence for a predictive effect, or (iii) enter an "expansion phase", where additional cohort of patients are recruited to improve the power for final analysis. The proposed framework addresses key methodological challenges, including interim decision rules, determining the magnitude and allocation of post-interim sample size increases, controlling the Type I error rate, and achieving meaningful gains in statistical power at the final analysis. The impact of an unknown biomarker distribution is also considered. These challenges are explored through simulation studies, where properties of the interim and final analyses are presented under various design configurations (where the decision bounds and size of the expansion are altered) and scenarios, in which the true predictive and prognostic effects of the biomarker are varied. Evaluated metrics include statistical power, Type I error control, and interim decision characteristics. The results inform the design of trials incorporating interim analyses to assess the predictive effect of a biomarker, enabling statistically efficient and well-calibrated decision-making.