Author(s)

Ruoyu Wang

Corresponding Author

Ruoyu Wang

Abstract

Vaccines are still the strongest defense against viral infections. However, finding effective antigens through traditional methods takes time and heavy experimental effort. Many computational approaches try to speed up this process, but most depend only on sequence data. They often overlook the structural and biochemical relationships that are key to understanding immunogenicity. In this study, we introduce a transformer-based multi-modal model for antigen prediction. The model combines three sources of information: physicochemical properties, amino acid sequences, and 3D structural features. A cross-attention fusion module connects these features and allows the model to learn how they interact. This design helps the system capture complex biological signals and improves recognition accuracy. We tested the model on bacterial and viral antigen datasets. It consistently performed better than traditional machine learning and single-modality deep learning methods. Across all metrics—accuracy, F1-score, and AUC—it showed strong improvements. The model can also locate highly immunogenic, surface-exposed fragments. Overall, it offers a clear, interpretable, and efficient computational tool for vaccine target discovery.

Keywords

Antigen prediction; Vaccine design; Multi-modal learning; Transformer; Cross-attention fusion