Vardan Harutyunyan, Mher Matevosyan, Andrea Brancale, Hovakim Zakaryan
Artificial intelligence (AI) is increasingly discussed across biomedical research, yet its adoption in antiviral drug discovery remains modest. This is paradoxical because antiviral research urgently needs precisely what AI can provide: faster target prioritization, more effective reuse of fragmented datasets, rapid drug repurposing, de novo molecular design and early resistance prediction. Limited adoption, however, does not simply reflect conservatism. Antiviral datasets are frequently small, heterogeneous, and assay-dependent; many models are difficult to interpret; and a substantial part of the literature still relies on docking scores or retrospective benchmarks rather than prospective biological validation. Skepticism among virologists, medicinal chemists, and pharmacologists is therefore understandable, but it should be directed at poorly designed workflows rather than at AI methods themselves. These limitations are not uniform across methods: supervised activity models are constrained mainly by data, generative models by how their objectives are specified, and general-purpose foundation models by domain mismatch and evaluation design, so each requires a different remedy. The field can advance by building curated antiviral datasets with standardized metadata, deploying interpretable and uncertainty-aware models, filtering predictions by pharmacological feasibility, and embedding AI within iterative computation-experiment loops. Sustained investment in training schools, hands-on workshops, and shared benchmark challenges will be essential if AI is to become a routine and trusted component of antiviral discovery.