Combination therapies play a crucial role in the treatment of complex diseases, such as cancer. They enhance efficacy, minimize resistance, and reduce toxicity by leveraging synergistic effects. However, identifying effective combinations is challenging due to the vast number of possible pairings and the high-priced costs of experimental validation. Machine learning (ML) and deep learning (DL) models have advanced drug synergy prediction by integrating diverse datasets and modeling the interactions between drugs and cell lines. Despite these advancements, most algorithms primarily rely on drug-specific features, such as chemical structures, with limited incorporation of functional drug information and cellular content features. We propose a novel approach that integrates Drug Resistance Signatures (DRS) as a biologically informed representation of drug information. This approach provides a more comprehensive framework for identifying effective combination therapies. We evaluated the predictive power of DRS features across various machine learning models (LASSO, Random Forest, AdaBoost, and XGBoost) and the deep learning model SynergyX. We compared their performance with that of conventional drug signatures and chemical structure-based descriptors. Our results demonstrate that models incorporating DRS features consistently outperform traditional approaches across all evaluated algorithms. Validation on independent datasets, including ALMANAC, O'Neil, OncologyScreen, and DrugCombDB, confirms the robustness and generalizability of the proposed framework. These findings emphasize the importance of integrating resistance-informed transcriptomic features into computational models. By capturing drug functionality in a biologically relevant context, DRS improves both the accuracy and interpretability of drug synergy prediction, offering a powerful strategy for guiding the discovery of effective combination therapies.
Mozaffarilegha et al. (Thu,) studied this question.