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

The COVID-19 pandemic has demanded urgent and accelerated action toward developing effective therapeutic strategies. Drug repurposing models (in silico) are in high demand and require accurate and reliable molecular interaction data. While experimentally verified viral-host interaction data (SARS-CoV-2-human interactions published on April 30, 2020) provide an invaluable resource, these datasets include only a limited number of high-confidence interactions. Here, we extend these resources using a deep learning-based multiview graph neural network approach, coupled with optimal transport-based integration. Our comprehensive validation strategy confirms 472 high-confidence predicted interactions between 280 host proteins and 27 SARS-CoV-2 proteins. The proposed model demonstrates robust predictive performance, achieving ROC-AUC scores of 85.9% (PPI network), 83.5% (GO similarity network), and 83.1% (sequence similarity network), with corresponding average precision scores of 86.4%, 82.8%, and 82.3% on independent test sets. Comparative evaluation shows that our multiview approach consistently outperforms conventional single-view and baseline graph learning methods. The model combines features derived from protein sequences, gene ontology terms, and physical interaction information to improve interaction prediction. Furthermore, we systematically map the predicted host factors to FDA-approved drugs and identify several candidates, including lenalidomide and pirfenidone, which have established or emerging roles in COVID-19 therapy. Overall, our framework provides comprehensive and accurate predictions of SARS-CoV-2-host protein interactions and represents a valuable resource for drug repurposing efforts.