A unified graph-based approach for protein function prediction using AlphaFold structures and sequence features

Abstract

Predicting protein function is a key challenge in computational biology with broad implications for understanding biological systems and disease mechanisms. Traditional deep learning approaches rely heavily on protein sequence data and protein–protein interaction (PPI) networks, often neglecting structural information due to limited availability of experimentally resolved protein structures. The advent of AlphaFold, which predicts protein structures with near-atomic accuracy, provides an opportunity to integrate structural context into function prediction. In this study, we propose StructSeq2GO, a novel hybrid model that combines structural and sequence information. StructSeq2GO employs graph representation learning to extract structural features from AlphaFold-predicted protein structures and integrates them with sequence embeddings derived from the ProteinBERT language model to predict Gene Ontology (GO) labels. Experimental evaluations demonstrate that StructSeq2GO achieves state-of-the-art performance across three GO domains, with $F_{max}$ scores of 0.485, 0.681, and 0.663, AUC scores of 0.764, 0.939, and 0.891, and AUPR scores of 0.688, 0.763, and 0.702 for the Biological Process (BPO), Cellular Component (CCO), and Molecular Function (MFO) ontologies, respectively. These results highlight the critical importance of structural information and the efficacy of ProteinBERT in enhancing protein function prediction, as structure provides spatial and biochemical context not captured by sequence alone. The model’s performance is influenced by the quality of AlphaFold structural predictions and may benefit from future improvements in structure confidence modeling. Additionally, extending StructSeq2GO to predict pathway-level or disease-related annotations could broaden its biological utility.