Experimental protein function annotation does not scale with the fast-growing sequence databases. Only a tiny fraction (<0.1%) of protein sequences has experimentally determined functional annotations. Computational methods may predict protein function very quickly, but their accuracy is not very satisfactory. Based upon recent breakthroughs in protein structure prediction and protein language models, we develop GAT-GO, a graph attention network (GAT) method that may substantially improve protein function prediction by leveraging predicted structure information and protein sequence embedding. Our experimental results show that GAT-GO greatly outperforms the latest sequence- and structure-based deep learning methods. On the PDB-mmseqs testset where the train and test proteins share <15% sequence identity, our GAT-GO yields Fmax (maximum F-score) 0.508, 0.416, 0.501, and area under the precision-recall curve (AUPRC) 0.427, 0.253, 0.411 for the MFO, BPO, CCO ontology domains, respectively, much better than the homology-based method BLAST (Fmax 0.117, 0.121, 0.207 and AUPRC 0.120, 0.120, 0.163) that does not use any structure information. On the PDB-cdhit testset where the training and test proteins are more similar, although using predicted structure information, our GAT-GO obtains Fmax 0.637, 0.501, 0.542 for the MFO, BPO, CCO ontology domains, respectively, and AUPRC 0.662, 0.384, 0.481, significantly exceeding the just-published method DeepFRI that uses experimental structures, which has Fmax 0.542, 0.425, 0.424 and AUPRC only 0.313, 0.159, 0.193.
