Drug combinations play an important role in therapeutics due to its better efficacy and reduced toxicity. Since validating drug combinations via direct screening is prohibitively expensive due to combinatorial explosion, recent approaches have applied machine learning to identify synergistic combinations for cancer. However, these approaches is not readily applicable to many diseases with limited combination data. Motivated by the fact that drug synergy is closely tied with biological targets, we propose a model that jointly learns drug-target interaction and drug synergy. The model, parametrized as a graph convolutional network, consists of two parts: a drug-target interaction and target-disease association module. These modules are trained together on drug combination screen as well as abundant drug-target interaction data. Our model is trained and evaluated on two SARS-CoV-2 drug combination screens and achieves 0.777 test AUC, which is 10% higher than the model trained without drug-target interaction.