DeepMolVir: Learning Virtual Virus Representations for Sequence-Aware Virus–Inhibitor Activity Prediction

The growing threat of emerging and drug-resistant viruses demands efficient approaches for predicting virus–inhibitor interactions. However, most existing computational models treat viruses as categorical identifiers and fail to capture sequence-level biological information. Here, we assembled a data set of 50,711 compound–virus activity pairs spanning 178 viral species and explored protein language models as a strategy for viral representation learning. Using embeddings generated by ESMC-600M, we introduce a virtual virus representation in which viruses are encoded as sequence-derived biological entities rather than as taxonomic labels. These viral representations are integrated with molecular features to model compound–virus interactions. In extensive 5-fold cross-validation, DeepMolVir demonstrated strong predictive performance, achieving an AUC of 0.8653 and an RMSE of 0.9480. Notably, virtual virus representations consistently outperformed conventional taxonomy-based encodings, while ablation studies highlighted viral sequence information as a key contributor to the model performance. Case studies on Nipah virus and hantaviruses further demonstrated the ability of DeepMolVir to identify candidate antiviral scaffolds and support cross-family drug repurposing. Collectively, our findings establish protein-language-model-based viral representation learning as an effective strategy for sequence-aware antiviral activity prediction and provide a foundation for computational antiviral discovery. Code and data are available at GitHub: https://github.com/Sunny-RPX/DeepMolVir.