Predicting functions of maize proteins using graph convolutional network.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Maize is a crucial global crop and model organism for gene function research.
• Gene Ontology (GO) annotation of maize proteins is challenging due to its large, hierarchical structure.
• Existing deep learning models for protein function prediction show limitations in utilizing GO hierarchy.

Purpose of the Study:

• To develop an accurate deep learning model for predicting Gene Ontology (GO) annotations of maize proteins.
• To effectively leverage the hierarchical structure of GO in protein function prediction.
• To improve the accuracy of protein function annotation by integrating sequence and GO hierarchy information.

Main Methods:

• Proposed DeepGOA, a deep Graph Convolutional Network (GCN) based model.
• Quantified correlations between GO terms and updated edge weights using GO annotations and hierarchy.
• Applied GCN to learn semantic representations of GO terms and CNN for amino acid sequence features.
• Integrated GO semantic representations and amino acid sequence features for end-to-end training.

Main Results:

• DeepGOA effectively integrates GO structural information and amino acid sequence data.
• The model achieves accurate GO annotations for proteins.
• Experiments demonstrate DeepGOA's superior performance compared to state-of-the-art methods on Maize and Human datasets.
• Ablation studies confirm the effectiveness of GCN in utilizing GO knowledge.

Conclusions:

• DeepGOA outperforms existing deep learning methods for protein function prediction.
• The Graph Convolutional Network (GCN) component significantly boosts prediction performance by leveraging GO hierarchy.
• The developed model offers an accurate and effective solution for maize protein function annotation.