Mapping antigenic evolution of influenza A virus using deep learning-based prediction of hemagglutination inhibition titers

Seasonal influenza remains a significant public health challenge through unpredictable antigenic drift, where accumulated mutations enable immune evasion and necessitate vaccine updates. Comprehensive antigenic characterization has been hindered by the absence of real-world A(H1N1)pdm antigenic mapping and post-2012 A(H3N2) maps due to insufficient pair-wise hemagglutination inhibition (HAI) titrations. Here, we present an end-to-end transformer model that predicts HAI titers directly from viral genetic sequences with error under two-fold, comparable to experimental variability. This approach enables rapid and high-throughput augmentation of HAI titrations across viral isolates, allowing for constructing large-scale antigenic maps. Several novel evolutionary patterns emerged from our analysis. Our antigenic mapping identified three A(H3N2) clusters between 2012-2022 with transitions occurring approximately every 6 years. Notably, genetically diverse co-circulating subclades 3C.2a and 3C.3a (2015-2020) formed a single antigenic cluster. A(H1N1)pdm formed three less temporally distinct clusters, with a novel cluster emerging post-COVID-19. Using interpretable analysis, we identified mutation sites linked to antigenic cluster transitions that align with previous laboratory findings. Key A(H3N2) mutations primarily occurred within major antigenic epitopes, while A(H1N1) showed fewer key mutations, some outside recognized antigenic regions. Our deep learning approach accelerates antigenic characterization in surveillance at global scale and can be transferred to other variable pathogens, providing an actionable bridge between genomic sequencing and vaccine strain selection.