A Deep Learning Approach to Antigenic Modeling for Rapidly Mutating Viruses

The evaluation of antigenic similarity between strains of viruses is a crucial aspect of vaccine production. The conventional methodology employed to quantify this degree of similarity is based on the performance of immunological assays, which are labor- and time-intensive. In this paper, optimized machine and deep learning models were applied to predict antigenic variants one of a rapidly mutating virus, foot-and-mouth disease. A comparison of embedding methods is provided. Models were also trained using reduced amino acids alphabet encoding, which allows the generation of new feature sets. These feature sets reduce model complexity and increase ensemble diversity. Based on diversity metrics, a stacking ensemble was built. The evaluation of models was conducted on a large dataset to provide robust results.