HNetGO: protein function prediction via heterogeneous network transformer.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Protein function annotation is crucial for understanding molecular life post-genome.
• Integrating multisource data improves protein function prediction, but existing methods face challenges with feature engineering and model integration.
• Deep learning models often overlook unlabeled sequence data, limiting their feature extraction capabilities.

Purpose of the Study:

• To develop an end-to-end protein function annotation model, HNetGO.
• To leverage heterogeneous networks for integrating protein sequence similarity and protein-protein interaction data.
• To utilize pretraining models for extracting semantic features from protein sequences.

Main Methods:

• HNetGO employs a heterogeneous network to combine protein sequence similarity and protein-protein interaction information.
• A pretraining model is used to extract semantic features from protein sequences.
• An attention-based graph neural network extracts node-level features from the heterogeneous network for function prediction.

Main Results:

• HNetGO achieves state-of-the-art performance on the human dataset.
• The model demonstrates superior accuracy in predicting protein functions related to cellular components.
• Significant improvements were observed in predicting molecular functions.

Conclusions:

• HNetGO offers an effective end-to-end solution for protein function annotation.
• The integration of heterogeneous networks and pretraining models advances protein function prediction.
• The model shows strong potential for biological research, particularly in understanding cellular and molecular functions.