Abstract
The advancement of metagenomic sequencing has revealed a vast viral diversity while simultaneously exposing limitations of homology-based tools such as BLAST and HMMER, which often fail to detect highly divergent viral genomes. The integration of artificial intelligence (AI) into viromics has transformed this landscape, introducing machine learning and deep learning models-including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers-that extend viral discovery beyond sequence similarity constraints. Structure-based frameworks such as AlphaFold, ESMFold, and Foldseek further enable annotation of divergent viral proteins through conserved 3D folds, while graph neural networks (GNNs) model host-virus interaction and explainable AI enhances interpretability of prediction. Despite their high sensitivity and scalability, AI-driven approaches face notable challenges: computational burden, data set bias, limited explainability, and elevated false discovery rates. This review traces the evolution of computational virology from traditional methods to AI-based and hybrid frameworks. We examine landmark AI tools while underscoring the continuing importance of phylogenetics and functional annotation in contextualizing AI predictions. We propose an integrated workflow that combines AI pattern recognition with classical bioinformatics to enhance both scalability and interpretability. By addressing the limitations of solely AI-driven or traditional approaches, this review presents a unified computational strategy to accelerate viral discovery, enhance evolutionary insights, and strengthen global preparedness for emerging infectious diseases.