EL-DRP: An Evolutionary Learning-Based Multi-Omics Framework for Drug Response Prediction with Interpretable Biomarker Selection

Cancer is a complex disease driven by diverse genetic, epigenetic, and microenvironmental alterations that result in dysregulated cell proliferation and therapeutic resistance. Both inter- and intratumoral heterogeneity further complicate treatment outcomes, highlighting the critical need for effective and efficient drug response prediction in precision oncology. While deep learning models have yielded robust predictive performance, their limited interpretability remains a critical barrier to clinical translation. To overcome this, we propose EL-DRP, an evolutionary learning framework for drug response prediction that integrates multi-omics data, including mRNA expression, copy number variation, and single nucleotide polymorphisms. Central to EL-DRP is the use of an Inheritable Bi-objective Combinatorial Genetic Algorithm (IBCGA) for optimized feature selection within each omics modality. IBCGA identifies minimal yet informative subsets of features, which are subsequently integrated to construct a unified multi-omics predictive model. EL-DRP achieved an average AUROC of 0.791 ± 0.072 for 21 drugs using an average of 106 biomarkers per drug, most of which align with known drug mechanisms of action. For instance, among the camptothecin response–associated biomarkers, KDM4A and SCML2 are known to participate in camptothecin-induced DNA damage response, whereas SALL2, RBM10, BAHCC1, and PBRM1 regulate chromatin structure and genome maintenance, corroborating their mechanistic relevance. Importantly, this minimal biomarker panel alone retains robust predictive power with explainable biomarkers. These results provide a new perspective for discovering potential biomarkers associated with drug response prediction and lay the groundwork for future studies integrating clinical data with translational potential.