Protein-protein interaction (PPI) is the basis of the whole molecular mechanisms of living cells. Although traditional experiments are able to detect PPIs accurately, they often encounter high cost and require more time. As a result, computational methods have been used to predict PPIs to avoid these problems. Graph structure, as the important and pervasive data carriers, is considered as the most suitable structure to present biomedical entities and relationships. Although graph embedding is the most popular approach for graph representation learning, it usually suffers from high computational and space cost, especially in large-scale graphs. Therefore, developing a framework, which can accelerate graph embedding and improve the accuracy of embedding results, is important to large-scale PPIs prediction. In this paper, we propose a multi-level model LPPI to improve both the quality and speed of large-scale PPIs prediction. Firstly, protein basic information is collected as its attribute, including positional gene sets, motif gene sets, and immunological signatures. Secondly, we construct a weighted graph by using protein attributes to calculate node similarity. Then GraphZoom is used to accelerate the embedding process by reducing the size of the weighted graph. Next, graph embedding methods are used to learn graph topology features from the reconstructed graph. Finally, the linear Logistic Regression (LR) model is used to predict the probability of interactions of two proteins. LPPI achieved a high accuracy of 0.99997 and 0.9979 on the PPI network dataset and GraphSAGE-PPI dataset, respectively. Our further results show that the LPPI is promising for large-scale PPI prediction in both accuracy and efficiency, which is beneficial to other large-scale biomedical molecules interactions detection.