Bayesian optimization design for dose-finding based on toxicity and efficacy outcomes in phase I/II clinical trials

Area of Science:

• Clinical Pharmacology
• Biostatistics
• Drug Development

Background:

• Traditional phase I trials focus on maximum tolerated dose (MTD), assuming monotonic dose-response relationships.
• Biologic agents often exhibit non-monotonic (unimodal or flat) dose-efficacy curves, necessitating identification of an optimal dose balancing efficacy and toxicity.
• Existing optimal dose-finding designs often rely on parametric models with strong assumptions, which may not suit diverse drug actions.

Purpose of the Study:

• To propose a novel Bayesian optimization framework for identifying optimal doses of biologic agents in phase I/II clinical trials.
• To address the limitations of parametric models by employing nonparametric dose-response modeling.
• To evaluate the performance of the proposed design against existing methods through simulation.

Main Methods:

• The proposed design utilizes a Bayesian optimization framework with Gaussian process priors for nonparametric modeling of dose-response relationships.
• It explicitly incorporates uncertainty in dose selection.
• Operating characteristics were compared via simulations against Bayesian optimal interval, EffTox, and isotonic designs.

Main Results:

• The proposed Bayesian optimization design demonstrated strong performance in simulations.
• It provided more stable results compared to the compared designs.
• Accuracy in optimal dose estimation and the percentage of correct dose recommendations were superior.

Conclusions:

• The proposed nonparametric Bayesian optimization design is a robust and effective method for optimal dose finding of biologic agents in early phase trials.
• This approach offers improved accuracy and stability over existing methods, particularly when strong parametric assumptions are inappropriate.
• It provides a valuable tool for optimizing dose selection in the development of novel biologic therapies.