Generalized Likelihood Ratios for Designing Dose Optimization Studies of Targeted Therapies

Abstract

Dose optimization studies of new therapeutic agents aim to identify one or more promising doses for further evaluation in subsequent studies. Traditionally, dose optimization has focused on finding the maximum tolerated dose (MTD), assuming that drug activity and efficacy generally increase with increasing dose. For modern targeted agents, the dose-activity relationship is often non-monotone and such that activity starts to plateau or even decline before reaching the MTD. Finding the optimal biological dose (OBD) for a targeted agent requires considering both toxicity and activity in dose optimization. This article proposes a new design for finding the OBD that utilizes generalized likelihood ratios (GLRs) to measure statistical evidence regarding key scientific questions on toxicity and activity. This GLR-based design requires no parametric modeling assumptions and only assumes that the dose-toxicity relationship is monotone and that the dose-activity relationship follows a two-sided isotonic regression model. Compared with existing designs that operate under similar assumptions, the GLR-based design is more general and more flexible, and performs competitively in simulation experiments where drug activity starts to plateau or decline before reaching the MTD.