STAR-GO: improving protein function prediction by learning to hierarchically integrate ontology-informed semantic...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate protein function prediction is crucial for biological discovery but lags behind rapid sequence data growth.
• Computational methods use Gene Ontology (GO) terms for function annotation, but existing models struggle with evolving ontologies and unseen terms.
• Current models often focus on either GO term semantics or structure, limiting generalization capabilities.

Purpose of the Study:

• To develop a novel framework, STAR-GO, for enhanced zero-shot protein function prediction.
• To integrate both semantic and structural characteristics of GO terms for improved model adaptability.
• To address the limitations of existing models in handling evolving ontologies and unseen GO terms.

Main Methods:

• Developed STAR-GO, a Transformer-based framework for joint modeling of GO term semantics and structure.
• Integrated textual definitions and ontology graph structure to learn unified GO representations.
• Processed GO representations hierarchically and aligned them with protein sequence embeddings for function prediction.

Main Results:

• STAR-GO achieves state-of-the-art performance in protein function prediction.
• Demonstrated superior zero-shot generalization capabilities, particularly for unseen GO terms.
• Showcased the effectiveness of integrating GO term semantics and structure for robust predictions.

Conclusions:

• Integrating GO term semantics and structure provides a robust and adaptable approach to protein function prediction.
• STAR-GO offers a significant advancement in handling evolving biological ontologies and improving prediction accuracy.
• The framework provides a valuable tool for accelerating biological and therapeutic discovery through enhanced protein function annotation.