An Artificial Neural Network Based Ensemble Model for Predicting Antigenic Variants: Application of Reduced Amino Acid Alphabets and Word2Vec

Assessment of antigenic similarity between strains of the influenza virus is a crucial factor when planning vaccine compositions. To perform this, a gold-standard laboratory procedure, hemagglutination inhibition assay, is conventionally used. Despite its theoretical importance and accuracy, this procedure suffers from several shortcomings, including high time consumption. Therefore, various computer-aided and mathematical methods have been developed to acquire earlier knowledge on the antigenic characteristics of currently circulating viruses. In this paper, we introduce a state-of-the-art ensemble artificial neural network model based on features derived from multi-representation of antigenicity. Generally, each feature is generated from an optimized convolutional neural network whose input describes the genetic difference between viruses in a specific numerical space. The space is determined based on embedding of the genetic sequence by a reduced amino acid alphabet and the Word2Vec framework. Our experiments indicated that the proposed model outperformed approaches from the literature by achieving an accuracy level of 0.933 for the HINI subtype. This implies possible application of our model as a promising exploratory tool in practical tasks of virus control.