Drug synergy prediction plays a significant role in cancer combination therapy, effectively reducing experimental costs and accelerating the discovery of efficient drug combinations. However, existing computational methods for predicting drug synergy still face significant limitations. Most models rely on a single feature extraction pathway, failing to comprehensively capture the multimodal information in drug molecules. Moreover, they usually ignore the historical synergy patterns between drug pairs and the similarity in cell lines’ response patterns to drug combinations, thus failing to effectively use transferable synergy priors to improve model generalizability. Based on the limitations, a novel prediction model, DPSM-Synergy, is proposed. Through integrating a dual-path feature extraction and synergy matrix enhancement strategy, it significantly improves the accuracy of anticancer drug synergy prediction. DPSM-Synergy employs a dual-path architecture composed of a PubChem pretrained model and a graph neural network (GNN) to extract the chemical semantic features and molecular graph features of drugs and innovatively introduces drug synergy matrices and cell line synergy matrices to capture the historical synergy patterns of drug pairs and the response similarity of cell lines to drug combinations, respectively. Experimental findings derived from two benchmark data sets, DrugCombDB and OncologyScreen, indicate that DPSM-Synergy surpasses existing state-of-the-art methods across all evaluation metrics, improving AUC-ROC by 4.00% and 3.57% over the best baseline on the DrugCombDB and OncologyScreen data sets, respectively. This validates the efficacy of the dual-path feature extraction and the synergy matrix enhancement strategy.
Wang et al. (Tue,) studied this question.