RIMGOGraph: integrating AlphaFold-derived residue interaction graphs and protein language embeddings for...

Accurate protein function annotation remains a major challenge in computational biology because the growth of protein sequence and structural databases continues to outpace experimental characterization. Recent advances in structure prediction, particularly AlphaFold2, have created new opportunities to incorporate three-dimensional structural information into predictive frameworks. However, effectively integrating structural topology with sequence-derived biological signals remains an open problem. Here, we present RIMGOGraph, a multimodal deep learning framework that combines residue interaction graphs derived from predicted protein structures with contextual sequence embeddings generated by the pre-trained ESM-2 model. Structural relationships among residues are encoded using graph neural networks, while a co-attention fusion module adaptively integrates global sequence and structural representations to capture complementary functional signals. Using a curated human protein dataset with Gene Ontology annotations, RIMGOGraph demonstrated robust and competitive performance across Biological Process, Cellular Component, and Molecular Function prediction tasks, achieving favorable Fmax, AUC, and AUPR scores relative to representative baseline methods. Ablation analyses further supported the complementary value of structural modeling and cross-modal fusion. The framework showed particularly strong performance for Molecular Function and Cellular Component prediction, suggesting that structural context is especially informative for localized biochemical activity and subcellular localization. Overall, RIMGOGraph provides an effective and scalable framework for structure-guided protein function prediction in the post-AlphaFold era. Future extensions incorporating multi-species training, stricter low-homology benchmarking, and additional biological modalities may further enhance its generalizability and practical utility.