Dose-Finding Trial Designs for Combination Therapies in Oncology

Abstract

In the article that accompanies this editorial, Gandhi et al present the results of a phase I combination study of neratinib and temsirolimus in patients with advanced solid tumors. The primary objective was to estimate the toxicity contour of the combination and establish recommended phase II doses of this combination. The maximum-tolerated dose (MTD) combinations were chosen based on the highest tolerated dose of each agent achieving a targeted doselimiting toxicity (DLT) rate of less than one third. Twelve dose combinations were explored out of a possible 16 combinations, where the dose level for each cohort of two subjects was determined by a nonparametric up-and-down algorithmic-based sequential design. Two MTD combinations were identified: 200 mg neratinib/25 mg temsirolimus and 160 mg neratinib/50 mg temsirolimus. Historically, dose-seeking trials in oncology have geared toward establishing the MTD of a therapeutic regimen, with safety as the primary outcome. A fundamental assumption of these trials is that the dose toxicity and dose efficacy relationships are monotone increasing, whereby the highest dose found to be safe is also assumed to be the dose most likely to be efficacious. Several issues surrounding the design and identification of the MTD of a combination regimen need careful consideration. Ideally, an understanding of the underlying biologic rationale for the combination would be available, for example, are the toxicity profiles of the agents overlapping, or additive? Is the efficacy of the two agents’ additive, complementary, or synergistic? Typically, a set of predetermined dose-level combinations are explored based on the single-agent MTD or other preclinical data demonstrating synergy. The dose of one agent under investigation is escalated while the dose of the second agent remains constant until a tolerable combination dose level is achieved. Often all possible combination levels cannot be feasibly explored. Despite the increased testing of such combination treatments in oncology, few designs for dose escalation of two or more agents have been proposed. The nonparametric up-and-down design used in this phase I combination trial is a rule-based (algorithmic) design, with the fundamental assumption that for each of the two agents, the probability of toxicity of one agent is nondecreasing with dose when the dose of the other agent is fixed. Figure 1 depicts a schematic representation for the MTD determination for a single-agent dose-escalation trial using the traditional cohorts-of-3 design approach and the one-way isotonic regression approach. Since the observed toxicity rates may not always exhibit a monotonic dose-toxicity relationship, as in the simple illustrative Figure 1 example, a bivariate isotonic regression estimator is used at the end to smooth the observed toxicity surface of the drug combination. Compared to most rule-based designs, this design allows for the identification of a MTD contour for the combination as opposed to a single combination level as the MTD. As with any phase I design, the dose-escalation/de-escalation rules, the cohort size, and the target toxicity rates are all parameters that need to be established in this design before the start of the trial. The dose allocation for subsequent cohorts in rule-based designs such as this one is based on the observed DLTs and toxicities at the current dose combination. The dose escalation in this trial allowed for the escalation of only one agent at a time, with no skipping of dose levels. One additional element of this design requires the specification of the start-up rule or the run-in phase: how many patients to assign, at what starting dose level combinations, before reverting to the primary design? The authors carried out extensive simulations under different underlying dose toxicity profiles to study the design properties. In comparison to the traditional and a modified cohorts-of-3 design (which are both rule-based designs), the nonparametric up-and-down design had better operating characteristics in terms of the proportion of times the true MTD was recommended as the phase II dose, and the numbers of patients treated at or near the MTD (see supplementary material accompanying Gandhi et al). Model-based designs can be an attractive alternative to rulebased designs when: (1) number of dose levels to be explored for escalation/de-escalation is large (12 dose-level combinations in this case), and (2) agent(s) being tested are expected to have unknown dose efficacy outcomes (ie, highest dose not necessarily the most efficacious). In the first case, the rule based designs would typically require a larger number of patients to be treated if indeed the MTD is near the highest dose level. In the second case, the dose allocation decisions are based not only on safety but also a measure of efficacy that is quick and reliable to assess. The continual reassessment method (CRM) first introduced the concept of dose-toxicity models to guide the dose-finding process. The dose-toxicity model represents the investigator’s a priori belief in the likelihood of DLT according to delivered dose, which thereafter is updated sequentially using cumulative patient toxicity data. While the choice of the prior distribution is often debated in the Bayesian framework, CRM designs have proven to be robust to model misspecification as long as the models themselves are selected based on clinical knowledge. Model-based designs have demonstrated superior operating characteristics compared to JOURNAL OF CLINICAL ONCOLOGY E D I T O R I A L VOLUME 32 NUMBER 2 JANUARY 1