EPGAT: Gene Essentiality Prediction With Graph Attention Networks

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Identifying essential genes and proteins is crucial for understanding human biology and disease.
• Machine learning (ML) methods, particularly those using protein-protein interaction (PPI) networks, have been explored for essential gene prediction.
• Existing methods face limitations due to network centralities not being exclusive proxies of essentiality and traditional ML's inability to process graph-structured data.

Purpose of the Study:

• To develop a novel computational approach for accurate essential gene and protein prediction.
• To overcome the limitations of traditional network-based and ML methods in predicting gene essentiality.
• To leverage Graph Attention Networks (GATs) for learning essentiality patterns directly from PPI networks and multiomics data.

Main Methods:

• Proposed EPGAT, an Essentiality Prediction approach based on Graph Attention Networks (GATs).
• Integrated multiomics data as node attributes within the GAT framework operating on PPI networks.
• Benchmarked EPGAT across four organisms, including humans, for gene essentiality prediction.

Main Results:

• EPGAT achieved high accuracy in predicting gene essentiality, with ROC AUC scores ranging from 0.78 to 0.97.
• Significantly outperformed traditional network-based and shallow ML-based prediction methods.
• Demonstrated robust performance even with limited and imbalanced training datasets, outperforming node2vec.

Conclusions:

• EPGAT offers a powerful and effective computational method for identifying essential genes and proteins.
• The approach advances the field by effectively utilizing graph-based deep learning on integrated biological data.
• This work provides a valuable tool for biological and pathological research requiring precise essential gene identification.