Title: Predicting antigenicity of human influenza virus A (H3N2) using deep learning

Biography:

Abstract

The rapid and large-scale pathogenesis of influenza virus requires constant monitoring, and frequent vaccine development to protect the world population not only from seasonal influenza but also from novel influenza A viruses that could trigger a pandemic. Seasonal Influenza is an acute viral infection and is estimated to cause 3 to 5 million cases of severe illness and around 250,000 to 500,000 deaths worldwide1. Among the three subtypes, type A is the only one known to cause pandemics. Previously developed models utilize a wide range of predictive algorithms to model the antigenic distance of influenza a viruses and achieved great success. However, these models only measure the contribution of chosen amino acids as individuals, which lacks the context that changes of amino acids in Hemagglutinin may have composite effects since they form a 3D structure in space. Besides reporting point mutations with their association of influenza epidemic, they also involve only a limited number of amino acid properties4. Understanding the combination effect of point mutations of influenza, A and expanding the number of amino acids in the analysis may better unveil the relationship between HA sequence and its antigenicity. In this study, we design a convoluted neural network (CNN) to model the patterns of HA protein sequence to analyze the patterns introduced by individual mutations and their associated and combination effect. Furthermore, we systematically analyze all available amino acid properties for the predictability of H3N2 antigenicity. Particle swam optimization algorithm is used to construct the structure of CNN. This approach produces good results with 10-fold cross validation of over 94% unbiased estimate, and blind prediction of 100% accuracy.