Keywords: early oncology, MTD, combination
With the recent advancement in the revolutionary immuno-oncology (I-O) therapies, it becomes more and more difficult to demonstrate that any single agent is better than the standard of care in the oncology therapeutic area. The next stage of early oncology development, therefore, is to focus on identifying the best combinations of established therapies with new agents to either overcome drug resistance or achieve synergic effects. Although in most cases the combination therapy is the interest, safety/efficacy profile of the new agent alone must also be explored. As a result, such trials designed to have mono-therapy and combo-therapy arms in parallel. Finding the maximum tolerable dose (MTD) based on dose limiting toxicity (DLT) is critical to early oncology development. Current oncology dose finding methods include 3+3 design, modified toxicity probability interval (mTPI), continual reassessment method (CRM), and Bayesian optimal interval design (BOIN). MTD is usually estimated by isotonic regression with a pool-adjacent-violators algorithm (PAVA). However, none of these methods handles the correlation and interplay between arms. E.g., DLT rate in combo-therapy arm must be no less than that in mono-therapy arm for the same dose level of the new drug. Ignoring such constraint may result in selecting MTD that is contradictory to common sense due to the relatively small size for phase I studies. Also, treating the two arms as independent during the trial, we may lose some useful information for the dose-escalation and MTD estimation. To overcome these issues, we propose a series of new strategies to improve the current approaches (mTPI and BOIN), for mono-therapy and combo-therapy parallel dose-finding trials, so that information may be “borrowed” from each arm for more efficient and ethical dose escalation decisions. To enhance the probability of selecting the true MTD and to avoid the selection of contradictory MTDs in mono-therapy and combo-therapy, we also propose to modify the standard PAVA and to use two-dimensional PAVA (2D-PAVA). The performance of the proposed approaches will be evaluated based on the simulated data under different scenarios.