HGCLSynergy: Heterogeneous Graph Contrastive Learning for Drug Synergy Prediction

In recent years, drug combination therapy has become a common cancer treatment strategy in clinical medicine, reducing the risk of drug resistance and side effects compared to monotherapy. Here, we proposed a drug combination prediction model, HGCLSynergy, based on contrastive learning of heterogeneous graphs, which effectively integrated biological data of drugs and cell lines to improve the accuracy of anticancer drug interaction synergy prediction. Specifically, we constructed a heterogeneous information network and then utilized two view encoders, namely neighbor-based and metapath-based view encoders, for feature learning. The expressive power of node features was significantly enhanced through contrastive learning between the two views. Finally, the multilayer perceptron was applied to predict the drug combination synergistic. Experimental results on two benchmark datasets showed that HGCLSynergy outperformed six other classical methods. This demonstrated that HGCLSynergy was an efficient and reliable tool for predicting drug combinations.