A Generalized Model-Free Framework for Dose Optimization in Multiple-Dose Randomized Early-Phase Clinical Trials

The conventional "more-is-better" dose selection strategy, which centers around the maximum tolerated dose (MTD), has shown limitations in the context of novel agents such as targeted therapies and immunotherapies. Often, doses lower than MTD may exhibit comparable efficacy with fewer toxicity, and sometimes, reaching the MTD may not be feasible. The US Food and Drug Administration (FDA), through its Project Optimus, has introduced draft guidelines to reform the dose optimization and dose selection paradigm. A notable recommendation is the execution of a randomized, parallel dose-response trial following a dose-escalation study, aiming to evaluate and determine the optimal dose based on totality of data including safety and efficacy. We propose a comprehensive analysis and decision-making framework for multiple-dose randomized trials. Specifically, we introduce a generalized model-free framework (GMod-Free) that integrates the established dose-ranging trial methodology with novel dose-finding designs to achieve proof of concept (PoC) and dose optimization with limited sample sizes. A generalized likelihood ratio test was developed to establish PoC, employing both parametric and non-parametric bootstrap methods to determine the distribution of the test statistics under the null hypothesis. Upon establishing PoC, a curve-free model-averaging approach is then used to identify the optimal dose. Importantly, GMod-Free does not rely on any parametric model assumption about the dose-response relationships but leverages the possible dose-outcome order to borrow information across different doses, ensuring robust operating characteristics. Our simulation studies show that the proposed GMod-Free yields desirable performance in establishing PoC and selecting the optimal dose, suggesting that it is a more transparent and efficient methodology for optimizing doses for clinical trials.