SynergyGraph: predicting cell line specific drug combination synergy scores using knowledge graph representation and hypergraph modeling.

Identifying synergistic drug combinations is crucial for enhancing the effectiveness and safety of cancer treatments. However, many existing computational methods mainly depend on pairwise drug modeling or limited molecular descriptors, which overlook the complex biological relationships among drugs, cell lines, and target proteins. To address these challenges, we introduce SynergyGraph, a comprehensive framework that integrates various biomedical data through knowledge graph (KG) construction, embedding learning, and hypergraph-based modeling to predict drug synergy in a cell line-specific way. In the initial phase, a detailed KG is built by linking drugs, proteins, and cell lines through multiple biological associations. Entities and relations, along with their types, are embedded into a shared latent space using Word2Vec, while drug representations are further enriched with structural and physicochemical features. To capture higher-order biological interactions, we create a hypergraph where hyperedges connect known synergistic drug-drug-cell line triplets within a specified threshold, effectively reducing redundancy and improving safety in new combinations. We then employ UniGAT from the UniGNN framework to learn expressive node embeddings. A multi-module deep learning model comprising a BioEncoder, a UniGAT-based Graph Encoder, and a regression-based Decoder is trained to predict synergy scores for drug combinations. The SynergyGraph model demonstrates superior performance on a small subset of the DrugComb dataset. Our results show that SynergyGraph effectively captures complex biological interactions, overcomes pairwise limitations, and reduces redundant effects of dual-drug combinations.