An Effective GCN-based Hierarchical Multi-label classification for Protein Function Prediction

Abstract

We propose an effective method to improve Protein Function Prediction (PFP) utilizing hierarchical features of Gene Ontology (GO) terms. Our method consists of a language model for encoding the protein sequence and a Graph Convolutional Network (GCN) for representing GO terms. To reflect the hierarchical structure of GO to GCN, we employ node(GO term)-wise representations containing the whole hierarchical information. Our algorithm shows effectiveness in a large-scale graph by expanding the GO graph compared to previous models. Experimental results show that our method outperformed state-of-the-art PFP approaches.

GOProFormer: A Multi-Modal Transformer Method for Gene Ontology Protein Function Prediction

“…We first provide a detailed look into the GOProFormer performance and then compare it to DeepGoPlus [ 14 ], DeepChoi-SeqVec [ 17 ], and DeepChoi-ESM-1, where we replace the SeqVec language model with the more powerful ESM-1. Support for the AllenNLP library employed by DeepChoi for SeqVec has been discontinued, so we present a new variant of DeepChoi, DeepChoi-ESM-1 that utilizes the ESM-1 PLM also employed in our proposed GOProFormer.…”

“…The highest values in each category are highlighted in boldface font. If we rely on reported values on the TDNK dataset, DeepChoi-SeqVec reports test-set Fmax values of , , and AURPC values of , , and , respectively, on the MF, BP, and CC sub-ontologies in the original paper [ 17 ].…”

“…As DeepGO [ 13 ], DeepGOPlus [ 14 ], DeepGOA [ 15 ], DeepChoi [ 17 ], and others, GOProFormer utilizes the amino-acid sequence of a protein and learns an embedding of a given protein sequence in a D -dimensional space. Specifically, considering protein in a given dataset, and its sequence of amino acids, GOProFormer learns the corresponding representation ; in this representation, each amino acid of the protein is mapped into .…”

“…Though only available in pre-print, DeepChoi [ 17 ] follows the overall approach in the 2019 DeepGOA [ 15 ] of learning sequence- and GO term representations and then combining them via a dot-product. The main difference is that the sequence representation is learned via SeqVec.…”

“…However, PLMs have yet to be demonstrated useful for GO annotation prediction. DeepChoi [ 17 ], a method presented in an article that remains in preprint, is the only occurrence of a PLM-based GO annotation method.…”