A graph neural network-based approach for predicting SARS-CoV-2-human protein interactions from multiview data.

Area of Science:

• Computational biology
• Virology
• Drug discovery

Background:

• The COVID-19 pandemic necessitates rapid development of therapeutic strategies.
• Accurate molecular interaction data is crucial for in silico drug repurposing models.
• Existing SARS-CoV-2-human interaction datasets are limited in high-confidence interactions.

Purpose of the Study:

• To extend existing resources by predicting high-confidence SARS-CoV-2-human protein interactions.
• To develop and validate a deep learning-based multiview graph neural network approach for interaction prediction.
• To identify potential drug repurposing candidates for COVID-19 therapy.

Main Methods:

• Utilized a deep learning-based multiview graph neural network approach with optimal transport integration.
• Integrated features from protein sequences, gene ontology terms, and physical interaction data.
• Validated predictions through a comprehensive strategy and compared performance against baseline methods.

Main Results:

• Successfully predicted 472 high-confidence interactions between 280 host proteins and 27 SARS-CoV-2 proteins.
• Achieved robust predictive performance with ROC-AUC scores ranging from 83.1% to 85.9%.
• Identified lenalidomide and pirfenidone as potential drug repurposing candidates for COVID-19.

Conclusions:

• The developed multiview graph neural network framework provides accurate and comprehensive predictions of SARS-CoV-2-host protein interactions.
• The findings offer a valuable resource for accelerating drug repurposing efforts against COVID-19.
• The study highlights the potential of integrating diverse data sources for improved molecular interaction prediction.