Path-Based Graph Neural Network for Drug Synergy Prediction and Interpretation

Combination therapy is a method of treating complex diseases by using multiple drugs, which has the advantages of good efficacy and few toxic side effects. It has been widely used in clinical research. The significant increase in the number of drug combinations in recent years has made drug screening more challenging. Therefore, the calculation method for predicting drug synergy has received widespread attention and achieved a superior performance. But there are still challenges. Many methods fail to interpret the model behavior or answer questions about the mechanism of action between drug combinations and cell lines. To address these challenges, we propose a path-based interpretable graph neural network model to predict and interpret synergistic drug combinations, named SDCInterpreter. SDCInterpreter enables the generation of simulated interpretations of drug synergistic mechanisms by exploring the relationships among drugs, genes, pathways, and cell line biomedical entities. Specifically, SDCInterpreter constructs a heterogeneous graph and uses a relational graph convolutional network for node representation learning. Mask learning and Dijkstra's shortest path algorithm are then used to mine the edges that have the most impact on drug synergy prediction and generate mechanism paths for interpreting drug synergy prediction. Our experimental results show that SDCInterpreter performs well in predicting drug synergy and effectively interprets the synergistic prediction results.