Prioritizing Influenza Vaccine Selection with Sequence-Based AI Antigenicity Models

Antigenic drift in influenza viruses poses a significant challenge for vaccine development, including vaccines based on previously circulating strains. Antigenicity is primarily assessed with the hemagglutination inhibition (HAI) assay, which measures how effectively elicited antibodies block HA-mediated agglutination. We present a sequence-based AI model trained on human H3N2 and H1N1 HAI data collected up to 2015, designed to predict HAI outcomes between vaccine and target strain hemagglutinin (HA) proteins. By leveraging historical data, the model enables inference of the antigenicity of candidate vaccines against both existing and newly emerged viral strains. On post-2015 evaluations, the model achieves an AUROC of ~0.70 when predicting responses of new strains to existing vaccines, and an AUROC of ~0.70 for new vaccines against existing strains. Performance declines to AUROC ~0.60 when both vaccine and strain are new, highlighting the challenge of generalization in rapidly evolving viral landscapes. Our findings suggest that AI foundation models can improve the yield of HAI testing by prioritizing vaccine–strain pairs.