Hierarchical deep learning for predicting GO annotations by integrating protein knowledge.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Manual curation and experimental testing are precise but costly and slow for assigning Gene Ontology (GO) terms.
• High-throughput sequencing generates vast data, necessitating automated methods for protein function prediction.
• Existing deep learning models for protein function prediction primarily use sequence data and lack breakthrough performance.

Purpose of the Study:

• To introduce DeeProtGO, a novel deep learning model for predicting GO annotations.
• To integrate diverse protein knowledge for enhanced functional prediction accuracy.
• To address the challenges in automatic protein function annotation.

Main Methods:

• Developed DeeProtGO, a deep learning model integrating protein knowledge.
• Trained the model on 18 prediction tasks across GO sub-ontologies, protein types, and taxonomic kingdoms.
• Benchmarked DeeProtGO against state-of-the-art methods using public datasets.

Main Results:

• Prediction quality increased with the integration of more protein knowledge.
• DeeProtGO demonstrated improved prediction of GO annotations compared to existing methods.
• The model effectively enhances the accuracy of automatic protein function prediction.

Conclusions:

• Integrating multiple sources of protein knowledge significantly improves GO annotation prediction.
• DeeProtGO offers a powerful computational tool for advancing protein research.
• The study highlights the potential of deep learning in addressing large-scale functional genomics challenges.