Computational efficiency comparison.

Proteins are fundamental to the execution of biological activities, and the accurate prediction of their functions is of paramount importance for protein research. Recent advancements in deep learning, particularly those based on Graph Neural Networks (GNNs), have demonstrated promising results by integrating protein graph features with sequence information. However, traditional GNN methods exhibit limitations in their feature representation capabilities, failing to capture long-range dependencies within sequences and lacking incorporation of inter-annotation relationships. To address these challenges, we propose a method, MAGIN-GO, which combines Graph Isomorphism Network (GIN) and Graph Convolutional Network (GCN) with Graph Convolutional Self-Attention Network (GMSA) to extract multi-source protein information and integrates Gene Ontology (GO) annotation embeddings. Our method effectively combines protein sequence features with protein-protein interaction (PPI) graph node features, extracts topological and contextual information through GIN and GMSA, and integrates pre-trained GO term embeddings into a multi-label classification framework. Comprehensive experiments on the UniProtKB/Swiss-Prot dataset demonstrate that MAGIN-GO outperforms existing methods, achieving AUPR values of 0.569, 0.434, and 0.754 for Molecular Function (MF), Biological Process (BP), and Cellular Component (CC) domains, respectively, with corresponding Fmax scores of 0.568, 0.458, and 0.752, Smin scores of 11.297, 37.709, and 8.079, and AUC scores of 0.896, 0.897, and 0.940. The experimental results showed that the performance of MAGIN-GO was good and superior to the existing methods.