A Deep Learning Framework for Predicting Protein Functions With Co-Occurrence of GO Terms

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate protein function prediction is crucial for drug discovery and crop improvement.
• A significant gap exists between the rapid increase in protein sequence data and functional annotations.
• Current methods often overlook the valuable information present in the co-occurrence of Gene Ontology (GO) terms.

Purpose of the Study:

• To develop a novel deep learning model, DeepPFP-CO, for improved protein function prediction.
• To leverage the co-occurrence patterns of GO terms to enhance prediction accuracy.
• To address the limitations of existing methods that treat GO terms as independent labels.

Main Methods:

• Implementation of a deep learning model, DeepPFP-CO, utilizing Graph Convolutional Networks (GCN).
• Exploration and integration of GO term co-occurrence relationships within the GCN framework.
• Fusion of predicted central function propensities with co-occurring function information for comprehensive prediction.

Main Results:

• DeepPFP-CO demonstrated superior performance compared to state-of-the-art methods like DeepGOPlus and DeepGOA.
• Evaluation metrics, including Fmax and AUPR, confirmed the enhanced accuracy of DeepPFP-CO.
• Protein-level analysis validated the model's effectiveness in improving protein function prediction.

Conclusions:

• DeepPFP-CO effectively captures and utilizes GO term co-occurrence for more accurate protein function prediction.
• The model offers a significant advancement over existing methods, addressing the challenge of limited functional annotations.
• DeepPFP-CO provides a valuable tool for biological research, available at https://csuligroup.com/DeepPFP/.