GatedGeoGO:Multi-Modal Geometry-Aware Network with Gated Fusion and GO Semantic Attention for Protein Function Prediction

Proteins play essential roles in diverse biological processes, and accurate function annotation is fundamental for understanding cellular mechanisms and disease pathogenesis. However, existing protein function prediction methods often lack effective multimodal integration and fail to fully exploit the rich semantic information in Gene Ontology (GO), limiting their ability to generalize across diverse proteins. To address these challenges, we propose GatedGeoGO, a novel deep learning framework that integrates multisource knowledge, including protein sequences, three-dimensional structures, protein-protein interactions (PPIs), and GO semantics. GatedGeoGO employs a gated fusion mechanism to selectively retain informative PPI embeddings, a geometry-aware protein graph network to capture multiscale structural features, and a GO-guided cross-attention module to dynamically inject semantic information, enabling context-aware dynamic multimodal fusion. Extensive experiments on benchmark datasets demonstrate that GatedGeoGO significantly outperforms state-of-the-art methods, particularly in predicting low-frequency GO terms, highlighting the effectiveness of advanced fusion strategies for large-scale protein function prediction.