Inferring anti-viral drugs using nonnegative matrix factorization with self-paced learning and hypergraph regularization

Abstract

Viruses pose significant threats to global health, and the development of effective antiviral therapies is challenging because of the high costs and long timelines associated with drug development. Drug repurposing has emerged as a promising approach to accelerate treatment discovery. In this study, we introduce a novel framework for predicting virus–drug associations, termed Preprocessing–enhanced Self-Paced Learning and HyperGraph regularized Nonnegative Matrix Factorization (Pre–SPLHGNMF). The model integrates linear neighborhood similarity to enhance similarity fusion, and Gaussian processes to reduce the impact of false negatives. It employs hypergraph regularization to capture higher-order relationships in virus–drug, virus–virus, and drug–drug associations, improving the model’s ability to make reliable predictions. Self-paced learning further optimizes the model by iteratively introducing samples from simple to complex, thus enhancing robustness against noisy data, ensuring more accurate association predictions. The experimental results show that Pre-SPLHGNMF outperforms the other methods in 10-fold cross-validation, achieving AUCs of 0.8526 and 0.8629 on the HDVD and VDA datasets, respectively. These findings highlight the potential of Pre-SPLHGNMF to advance virus–drug association prediction and support the development of effective antiviral strategies.