Maximizing Efficiency of Artificial Intelligence‐Driven Drug Combination Optimization through Minimal Resolution Experimental Design

Abstract

AbstractCombination therapy is critical to treating complex diseases, such as cancer. Optimizing drug combinations can be challenging when dealing with limited resources, such as patient tumor biopsies. This is due to the difficulty in screening large numbers of drug‐dose combinations as well as heterogenous patient responses. Developing a data‐driven approach that can efficiently and accurately determine the best drug combinations with the smallest number of experiments will have broad applications in both drug development as well as precision medicine. Quadratic Phenotypic Optimization Platform (QPOP) is an artificial intelligence (AI) approach that utilizes small experimental drug response datasets to accurately identify globally optimal drug combinations. QPOP exploits orthogonal array composite design (OACD), a combination of two‐level fractional factorial and three‐level orthogonal array components, to generate a minimal set of drug combinations sufficient for second‐order model fitting that is necessary to derive optimal drug combinations. In this study, a more efficient resolution IV OACD is able to derive optimal drug combinations and parabolic response surface maps, despite requiring fewer experimental runs than resolution V. As such, this study provides the framework for a more efficient OACD that can be applied toward AI‐driven drug combination design in patient‐based drug development and precision medicine.