Table 1.
Summary of the UniProtKB/Swiss-Prot dataset.
Fig 1.
The framework of the MAGIN-GO.
(a) The GIN module learns local topological patterns in the PPI network to generate graph-enhanced features; (b) Sequence embedding features are fused with the PPI network to construct protein representations; (c) The GMSA module captures long-range dependencies and global contextual interactions within and between sequences, complementing the local network features from GIN; (d) Pre-trained GO term embeddings are fused with the dual-module output features; (e) The fused features enable protein functional annotation prediction through a residual classifier.
Fig 2.
The structure of the GMSA module involves processing the inputs through a GCN to generate the query vector (Q), key vector (K), and value vector (V). Protein context information is then captured using a multi-head self-attention mechanism.
Table 2.
Experimental results on UniProtKB/Swiss-Prot data - Part 1 (mean ± std).
Table 3.
Experimental results on UniProtKB/Swiss-Prot data - Part 2 (mean ± std).
Table 4.
Computational efficiency comparison.
Table 5.
Ablation experiment results on UniProtKB/Swiss-Prot data.
Table 6.
Ablation experiment results on UniProtKB/Swiss-Prot data.
Table 7.
Effectiveness of GIN experiment results on UniProtKB/Swiss-Prot data.
Fig 3.
Evaluation metrics comparison among different models.
Table 8.
Summary of the CAFA3 dataset.
Table 9.
Experiment results on CAFA3 data.
Fig 4.
GO term frequency on three ontologies.
Fig 5.
Virtual functional subgraph analysis on MF, BP, and CC.
Table 10.
Examples of yeast and human protein function prediction.