Sequence-structured fitness landscapes predict influenza evolution

Seasonal influenza viruses continually evolve to evade host immunity. Vaccines must be regularly updated to protect against new circulating strains, motivating efforts to forecast viral evolution. Recent experimental and computational methods have made progress on this problem, but learning the fitness landscape driving viral evolution remains challenging because genomic surveillance samples only a small fraction of the possible sequence space. Here, we developed a generative framework that shares information across genetically related variants, inferring effective mutation-level contributions to viral fitness from temporal sequence data. Applied to hemagglutinin and neuraminidase from A/H3N2 and A/H1N1 viruses, the inferred fitness landscapes recovers known determinants of antigenicity and reveals convergent spatial organization across subtypes consistent with common antibody-mediated selection. Forward simulations from these landscapes predicted both the fixation or loss of circulating mutations and the future composition of viral populations, outperforming models based on neutral evolution or multinomial-logistic growth. Collectively, our results show that interpretable representations of pathogen fitness landscapes can reveal biologically important features and improve prospective evolutionary forecasting.