Benefiting from the advancements in computational methods, drug-protein interactions (DPIs) prediction has garnered increasingly attention in drug development processes. However, existing DPIs prediction models still encounter challenges in efficiently extracting node features from complex networks. This paper proposed a novel DPIs prediction framework named GSDPI, in which graph neural networks (GNN) were employed to aggregate neighborhood information of complex heterogeneous networks and represent feature matrices of drugs and proteins. Then, singular value decomposition (SVD) technique was effectively applied to convert the feature matrices into compact representations. Finally, multiple rounds of bidirectional random walks were performed in the reconstructed network to predict novel DPIs. The results demonstrated GSDPI could gain better prediction performance than several state-of-the-art models, achieving prediction accuracies of 0.9840, 0.9846, 0.9767, and 0.9878 on four public datasets, respectively.