SEGT-GO: a graph transformer method based on PPI serialization and explanatory artificial intelligence for protein...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Protein function prediction is crucial but challenging due to vast sequence data.
• Protein-Protein Interaction (PPI) networks are vital for uncovering functional relationships.
• Current methods struggle with distant protein relationships in PPI networks, with diminishing gains from deeper graph networks.

Purpose of the Study:

• To develop a novel method for predicting protein function by effectively capturing distant functional relationships in PPI networks.
• To improve the accuracy and generalization of protein function prediction, especially for large-scale, multi-species datasets.

Main Methods:

• Proposed SEGT-GO, a Graph Transformer model utilizing multi-hop neighborhood serialization and Explainable AI (XAI).
• Multi-hop neighborhood serialization converts PPI network information into feature embeddings for deeper learning.
• SHAP XAI framework optimizes model input and reduces feature noise.

Main Results:

• SEGT-GO demonstrated competitive performance against DeepGraphGO on large datasets and superior results on small datasets.
• The method effectively extracts functional information from proteins at deeper network levels.
• Achieved superior results in cross-species learning and predicting functions of unseen proteins.

Conclusions:

• SEGT-GO shows strong generalization capabilities for protein function prediction.
• The approach effectively addresses the challenge of discerning functions from complex protein sequence data.
• Highlights the potential of combining graph transformers, serialization, and XAI for advanced bioinformatics tasks.