In drug discovery, machine learning approaches have been widely adopted for predicting compound protein interactions so as to reduce cost and accelerate drug screening process. In this paper, taking advantage of structural information of molecules, we demonstrated the effectiveness of graph convolutional neural network for predicting the interactions on a data set of compounds for target proteins binding to curcumin, which has been reported to suppress tumor progression in previous research. The results show that the graph convolutional neural network outperformed other conventional machine learning approaches such as random forest and support vector machine. In addition, we predicted interactions between curcumin derivatives and target proteins using the trained graph convolutional model. Furthermore, in order to interpret the prediction model and discover important features for the prediction, we produced molecular graphics highlighting important atoms and identified how much each atom contributed to the prediction.