UniRES-GO: Unified residue-level early fusion of sequence and predicted structure for protein function prediction

Protein function prediction remains a central problem in bioinformatics, with broad implications for understanding biological processes, disease mechanisms, and drug discovery. Due to the high cost and time required for experimental characterization, only a small fraction of proteins have reliable functional annotations, highlighting the need for accurate computational approaches. Recent advances in protein structure prediction, particularly AlphaFold2, have enabled large-scale access to high-quality three-dimensional structures, creating new opportunities for structure-informed function prediction. In this study, we propose UniRES-GO (Unified Residue-level Early Fusion for Gene Ontology prediction), a novel framework that integrates protein sequence features with AlphaFold2-predicted structural information via residue-level early fusion. The fused representations are modeled as protein contact graphs and processed using a Graph Attention Network to capture both local residue interactions and global structural context, yielding discriminative protein-level embeddings for multi-label function prediction. We evaluate UniRES-GO on a human protein dataset across the three Gene Ontology categories: Biological Process, Cellular Component, and Molecular Function. Experimental results demonstrate that UniRES-GO consistently outperforms representative sequence- and interaction-based methods across multiple evaluation metrics, including F_max, AUC, and AUPR. In particular, UniRES-GO achieves strong performance in Molecular Function prediction, reaching an AUC of 0.970, while maintaining high stability across multiple runs. Ablation studies further confirm the effectiveness of the residue-level fusion strategy and graph-based modeling. Overall, UniRES-GO provides an effective and generalizable approach for protein function prediction by leveraging predicted structural information, offering practical advantages for annotating proteins lacking homologous sequences or interaction data.