Bridging antiviral drug discovery with a large language model-powered framework

Viral infections pose ongoing threats to human health, emphasizing the continued need for effective antivirals. Antiviral drug discovery often relies on phenotype-based drug discovery (PBDD) and target-based drug discovery (TBDD). However, current computational approaches focus solely on predicting compounds that bind to specific antiviral-related targets, overlooking the biological relevance of antiviral phenotypes. Here, we propose DeepAVC, a large language model-powered framework that integrates DeepPAVC for PBDD and DeepTAVC for TBDD. As a result, DeepAVC outperforms existing baselines in antiviral compound prediction and provides high interpretability by identifying key atoms and residues involved in compound-protein interactions. Moreover, we demonstrate that DeepPAVC and DeepTAVC complement each other and can be used synergistically. We further confirm DeepAVC's power through both in vitro and in vivo experiments. Finally, we identify MNS as a novel broad-spectrum antiviral compound with greater efficacy than Sisunatovir. All these results suggest that DeepAVC is a valuable tool for antiviral drug discovery.