Using pathogen genomics to predict antigenic changes of influenza H3N2 virus

Influenza viruses evolve their antigenic phenotype in response to immune-driven selection, necessitating regular updates to the antigens used in influenza vaccines. Characterising antigenic change of influenza virus currently requires intensive serology-based laboratory work to analyse the movements of viruses in antigenic space. By contrast, pathogen genomics can generate large numbers of sequences in near real-time. However, using genetic sequences to predict antigenic changes remains challenging due to the complex relationship between genetic and antigenic changes. Here we develop a deep learning approach to predict the antigenic distance between influenza H3N2 viruses using 83,145 hemagglutinin (HA) sequences. We first train an elementary protein language model to encode the HA sequences, and then couple the trained language model with a deep regression model to realise the sequence-based prediction. Our method provides stable and accurate prediction on the antigenic distance between H3N2 viruses using their HA sequences. This facilitates the prediction of antigenic movements of viruses in each influenza season using their sequence data. Our sequence-based method will enable real-time prediction of influenza antigenic changes and support vaccine strain selection.