Predicting Protein Functions Based on Heterogeneous Graph Attention Technique.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Protein function prediction is vital for drug discovery and understanding disease mechanisms.
• Current methods often overlook negative experimental annotations, limiting prediction precision.
• Existing databases primarily contain positive protein function annotations.

Purpose of the Study:

• To develop an advanced deep learning method for accurate protein function prediction.
• To address the underestimation of precision caused by ignoring negative annotations.
• To improve the prediction of unobserved functional annotations.

Main Methods:

• Constructing a heterogeneous graph integrating protein-protein interactions, ontology structure, and negative annotations.
• Utilizing a heterogeneous graph attention technique to learn protein and ontology term embeddings.
• Reconstructing positive protein-term associations and scoring unobserved annotations.

Main Results:

• The proposed method demonstrated superior performance in predicting new protein annotations across Human, Mouse, and Arabidopsis.
• Incorporating limited negative annotations significantly enhanced predictive performance.
• The approach outperformed existing state-of-the-art methods in protein function prediction.

Conclusions:

• The novel deep learning method effectively leverages negative annotations for improved protein function prediction.
• This approach offers a significant advancement for bioinformatics research and applications.
• Accurate protein function prediction is crucial for biological discovery and therapeutic development.