The subcellular location of proteins is closely associated with their functions. Understanding the subcellular localization of proteins contributes to the comprehension of disease onset and progression. Although existing computational methods have demonstrated advantages over traditional approaches, the scarcity of labeled data remains a hindrance, especially for proteins with multiple locations. To address this issue, we propose a pre-training method for protein subcellular localization, named PrePSL, by using graph auto-encoder and protein language model. Firstly, we construct a network based on protein-protein interactions, employing a graph auto-encoder to capture intricate relationships among proteins. Next, a protein language model is adopted to extract feature from amino acid sequences, which provide important information about the subcellular localization of the protein. Finally, the fusion feature, an integration of two features acquired from both networks and sequences, is fed into a deep neural network (DNN) to predict subcellular locations of the proteins. In the five-fold cross-validation, PrePSL achieves optimal performance with RL, CV, and AP values of 0.0471, 0.9057, and 0.8537, respectively. Various ablation experiments further validate the model’s rationality, and case studies demonstrate the practical utility of the model. The source codes of PrePSL are available at: https://github.com/LabBioMedCoder/PrePSL .