Protein function prediction is essential for elucidating biological processes and accelerating drug discovery. However, the vast number of unannotated protein sequences and the limited availability of experimentally validated functional data remain major challenges. Although deep learning models based on protein sequences or protein-protein interaction networks have shown promise, their performance is still restricted, particularly for proteins without interaction data. Furthermore, many existing approaches treat sequence and structural information separately, potentially resulting in suboptimal feature representations. To address these limitations, we propose an improved graph-based framework that integrates two key innovations: (i) ESM-2, a state-of-the-art protein language model, to generate semantically rich sequence embeddings; and (ii) a hybrid pooling mechanism within graph convolutional blocks to better capture both global and local structural features from AlphaFold2-predicted structures. Experiments on the human proteome demonstrate that our model consistently outperforms existing methods in predicting molecular function, cellular component, and biological process annotations. These findings highlight the advantages of combining advanced sequence representations with enhanced structural learning for accurate and generalizable protein function prediction.
Protein function prediction has advanced substantially with the integration of AlphaFold2 structural information, yet current models remain constrained by incomplete multimodal feature fusion and limited attention mechanisms for capturing structural-functional relationships. Here, we present an enhanced framework that overcomes these limitations through three innovations: (i) Graph-CBAM, the first adaptation of convolutional block attention to graph neural networks for fine-grained structural feature extraction; (ii) complete multimodal fusion of Node2vec structural embeddings with amino acid one-hot encodings; and (iii) a dual-head self-attention pooling module that stabilizes node importance estimation. Extensive experiments on human protein data sets demonstrate that our model consistently outperforms existing benchmarks across all Gene Ontology branches. We report pronounced improvements, including an increase in Fmax by 2.9% on the Biological Process (BP) branch (0.481 to 0.495) and an enhancement of AUPR by 3.9% on the Cellular Component (CC) branch (0.763 to 0.793). Performance for Molecular Function (MF) remains competitive. Ablation analyses further confirm the independent contributions of structural embeddings, one-hot encodings, and Graph-CBAM. Overall, this work provides a more complete and practical solution for AlphaFold2-based protein function prediction, with particular advantages in predicting functions of proteins lacking protein-protein interaction data.