Seasonal H3N2 influenza evolves rapidly, leading to an extremely poor vaccine efficacy. Substitutions employed during vaccine production using embryonated eggs (i.e., egg passage adaptation) contribute to the poor vaccine efficacy (VE), but the evolutionary mechanism remains elusive. Using an unprecedented number of hemagglutinin sequences (n = 89,853), we found that the fitness landscape of passage adaptation is dominated by pervasive epistasis between two leading residues (186 and 194) and multiple other positions. Convergent evolutionary paths driven by strong epistasis explain most of the variation in VE, which has resulted in extremely poor vaccines for the past decade. Leveraging the unique fitness landscape, we developed a novel machine learning model that can predict egg passage substitutions for any candidate vaccine strain before the passage experiment, providing a unique opportunity for the selection of optimal vaccine viruses. Our study presents one of the most comprehensive characterizations of the fitness landscape of a virus and demonstrates that evolutionary trajectories can be harnessed for improved influenza vaccines.

Seasonal Influenza H3N2 virus poses a great threat to public health, but its vaccine efficacy remains suboptimal. One critical step in influenza vaccine production is the viral passage in embryonated eggs. Recently, the strength of egg passage adaptation was found to be rapidly increasing with time driven by convergent evolution at a set of functionally important codons in the hemagglutinin (HA1). In this study, we aim to take advantage of the negative correlation between egg passage adaptation and vaccine effectiveness (VE) and develop a computational tool for selecting the best candidate vaccine virus (CVV) for vaccine production. Using a probabilistic approach known as mutational mapping, we characterized the pattern of sequence evolution driven by egg passage adaptation and developed a new metric known as the adaptive distance (AD) which measures the overall strength of egg passage adaptation. We found that AD is negatively correlated with the influenza H3N2 vaccine effectiveness (VE) and ~75% of the variability in VE can be explained by AD. Based on these findings, we developed a computational package that can Measure the Adaptive Distance and predict vaccine Effectiveness (MADE). MADE provides a powerful tool for the community to calibrate the effect of egg passage adaptation and select more reliable strains with minimum egg-passaged changes as the seasonal A/H3N2 influenza vaccine.