Multi-View Fusion with Knowledge Graph for Anticancer Drug Synergy Prediction

Synergistic drug combinations inhibit cancer cells through multiple mechanisms of action, significantly improving cancer treatment efficacy and alleviating drug resistance. This makes them an effective approach for cancer therapy. In recent years, the drug combination prediction methods based on machine learning and deep learning have been developed to pre-screen potential synergistic drug combinations. Among these, some knowledge graph (KG)-based methods predict drug combinations by leveraging the rich entity relationships in the KG. However, existing KG-based methods either fail to fully exploit the interaction between drugs and cell lines or overlook the implicit knowledge embedded in semantically related but topologically disconnected entities in the KG. Additionally, knowledge graphs often introduce noise into the representation learning process due to the imbalance in reality data distribution. To address these, we propose MFKGSynergy, a drug synergy prediction model based on knowledge graph and multi-view information fusion. The model constructs three views: (i) Drugcell Line interaction view for global interaction information between drugs and cell lines, using a Graph Transformer attention mechanism; (ii) Knowledge Graph View: a heterogeneous graph relation-aware attention network is utilized to learn high-order topological information between drugs/cell lines and entities along connected paths in KG; (iii) Latent Semantic View, which captures implicit knowledge in the KG by calculating the similarity between drugs. Then, contrastive learning is employed to align drug and cell line representations across different views. And a gated aggregation module is designed to weight and fuse interaction and KG information, effectively reducing noise caused by data distribution imbalance. It is shown that our model improves prediction performance compared to the state-of-theart model.