Graph-based Drug Decomposition for Anticancer Response Modeling

Abstract This paper studies the molecular effects on ex vivo drug response in pediatric acute myeloid leukemia (AML). We firstly estimate dose–response relationships through linear and mixed-effects models, capturing both patient-specific heterogeneity and drug-level effects. Then, drug identifiers are decomposed into curated molecular descriptors and their higher-order interactions, yielding a structured, interpretable representation of chemical properties. To handle the resulting high-dimensional system, we introduce a specialized graph-based drug decomposition and selection, enabling a computationally tractable estimation of the molecular effects on ex vivo drug response. This strategy uncovers the molecular features most strongly associated with cellular viability, providing a biologically grounded and transparent alternative to black-box predictive methods. By directly linking molecular structure to therapeutic outcomes, our framework supports a novel mathematical-programming-based drug-combination selection and cost-efficient compound prioritization in pre-clinical leukemia research.