Bidirectional Semantic Proximity for Protein-GO Association on Heterogeneous Biological Networks

\noindent \textbf{Background} Protein function annotation is a fundamental task in computational biology, and accurate prediction of Gene Ontology (GO) terms for proteins is of great significance for understanding biological mechanisms and disease pathogenesis. However, existing computational methods are largely confined to single protein sequence features or the topological structure of protein-protein interaction (PPI) networks, failing to fully exploit the semantic associations within the GO hierarchy or the interactions among multi-source information in heterogeneous networks. This limitation leads to suboptimal prediction accuracy and generalization capability.\noindent \textbf{Methods} To address these challenges, we propose BSP (Bidirectional Semantic Proximity), a protein function prediction method based on a directed heterogeneous network. Specifically, we first construct a directed heterogeneous network comprising protein nodes and GO term nodes, systematically integrating three types of biological relationships: bidirectional PPI edges, directed hierarchical edges among GO terms, and protein--term reference annotation edges. On this basis, we design a bidirectional symmetric semantic proximity scoring function, which achieves bidirectional propagation and symmetric measurement of functional signals through PPI neighbor semantic aggregation on the protein side and GO ancestor semantic aggregation on the term side, thereby comprehensively characterizing the strength of functional associations between proteins and GO terms.\noindent \textbf{Results} Extensive experiments on four representative eukaryotic species datasets, namely Yeast, Human, Arabidopsis, and Fly, demonstrate that BSP achieves superior performance across all three GO sub-ontologies, namely Biological Process (BP), Cellular Component (CC), and Molecular Function (MF), significantly outperforming baseline methods such as FSWeight and TAFS. Ablation studies and robustness analyses further validate the effectiveness and stability of the directed graph architecture, loss function hyperparameter settings, and the multi-source information fusion mechanism. This study provides a systematic heterogeneous network modeling framework for protein function prediction, exhibiting promising cross-species generalization capability and practical application value.