Innovative Statistical Methods and Strategies for Dose Optimization in 21st Century Drug Development

There is an increasing interest in the development of the combinations of several drugs that can enhance the efficacy of each of the compounds compared to when administered as monotherapies. These could be combinations of a new experimental agent with an approved (or at least well-studied) compound or a combination of two new experimental agents. In either case, the selection of the optimal dose level of one agent might depend on the dose level of another agent. This implies that these two should be optimised together. Furthermore, the toxicity and efficacy of the combination might be different when its is administered under different schedules. In total, the doses of two agents and their administration schedule should be administered simultaneously resulting in a 3-dimensional dose-escalation problem with potentially many combination-schedule levels to be tried. Due to these challenges, it is common to restrict the search with at least one compound being fixed. This, however, can result in suboptimal selection of the combination-schedule. In this talk, motivating by a real trial, we will introduce a novel dose-escalation design that partitions the 3-dimensional grid of combination-schedules into smaller sub-grids and fits a combination-schedule-toxicity model within each. Then, via model-averaging between these models, the estimates for the whole 3-dimensional grid are obtained. We will show that under a number of considered cases, the proposed design results in better accuracy than other model-based designs that attempt to model the whole grid altogether.  

Project Optimus, an initiative by the FDA's Oncology Center of Excellence, seeks to reform the dose-optimization and dose-selection paradigm in oncology by integrating more data sources in the estimation of optimal doses. Most designs that account for patient heterogeneity are intended for trials where heterogeneity is known and pre-defined subpopulations are specified. However, given the limited information at such an early stage, subpopulations should be learned through the design. We propose a dose-optimization design that integrates toxicity, pharmacokinetic, patient characteristic, and response data to inform dose recommendations. The dose-optimization design is carried out in two stages. First, a toxicity-driven stage estimates a safe set of doses. Then, a dose-ranging efficacy-driven stage explores the set using response and patient characteristic data by employing Bayesian sparse group selection to understand patient heterogeneity. An optimal dose is recommended for each identified subpopulation within the target population. The simulation studies show that a model-based approach to identifying the target population can be effective; patient characteristics relating to heterogeneity were identified correctly and different optimal doses were recommended for each identified target subpopulation. Our findings show that accounting for heterogeneity is advantageous even in the more realistic setting of not knowing the source of heterogeneity. 

In recent years, dose optimization has emerged as a crucial element in the development of oncology drugs, with regulatory authorities increasingly mandating comprehensive data and justification for dosing before Phase 3 trials. Despite this, the practical implementation of optimization studies, especially in terms of design, remains a complex challenge. This presentation will delve into the intricacies of planning these studies, focusing on strategies for selecting appropriate sample sizes, evaluating alternative study designs, and making informed dose selection decisions.  

Detecting the efficacy signal and determining the optimal dose are critical steps to increase the probability of success and expedite the drug development in cancer treatment. After identifying a safe dose range through phase I studies, conducting a multi-dose randomized trial becomes an effective approach to achieve this objective. However, there have been limited formal statistical designs for such multi-dose trials, and dose selection in practice is often ad hoc, relying on descriptive statistics. We propose a Bayesian optimal two-stage design to facilitate rigorous dose monitoring and optimization. Utilizing a flexible Bayesian dynamic linear model for the dose-response relationship, we employ dual criteria to assess dose admissibility and desirability. Additionally, we introduce a triple-outcome trial decision procedure to consider dose selection beyond clinical factors. Under the proposed model and decision rules, we develop a systematic calibration algorithm to determine the sample size and Bayesian posterior probability cutoffs to optimize specific design operating characteristics. Furthermore, we demonstrate how to concurrently assess toxicity and efficacy within the proposed framework using a utility-based risk-benefit trade-off. To validate the effectiveness of our design, we conduct extensive simulation studies across a variety of scenarios, demonstrating its robust operating characteristics.