Abstract:
Precise prediction of antigenic proteins (APs) is crucial for accelerating vaccine development, antibody engineering, and immunotherapeutic research. This study presents a novel and effective deep learning framework the Multiheaded Residual Convolutional Neural Network, specifically designed to predict antigenicity directly from amino acid sequences with high specificity, accuracy, and robustness. Traditional experimental methods for antigenic protein identification, though accurate, are time-consuming, expensive, and labor-intensive, making them unsuitable for large-scale screening. To overcome these challenges, this research proposes a computationally efficient approach capable of learning deep hierarchical patterns and contextual dependencies within protein sequences, thereby enabling faster and more precise prediction of antigenic candidates. The dataset used for both training and testing was manually curated from the UniProt database and comprised experimentally validated antigenic and nonantigenic proteins. Protein sequences were encoded using two powerful feature extraction techniques ProtBERT and UniRep which effectively captured evolutionary, contextual, and structural relationships among amino acid residues. The proposed model achieved outstanding results, with 97% accuracy on the training dataset and 95% accuracy on the testing dataset, supported by strong sensitivity, specificity, and MCC values. The model’s reliability and generalization capability were further validated using 5-fold cross-validation, confirming its stability across diverse protein samples. This research demonstrates that the Multiheaded Residual Convolutional Nueral Network based computational model offers a scalable and biologically meaningful alternative to conventional laboratory methods. Its superior predictive capability not only enhances the efficiency of antigen identification but also paves the way for rapid advancements in computational immunology, vaccine design, and epitope-based therapeutic discovery.
