Predictability of the antigenic evolution of human influenza A H3 viruses

Abstract

The current influenza A antigenic evolution paradigm suggesting that antigenic evolution is highly constrained, with successful new viruses being near optimal at maximizing their antigenic distance from past strains. This begs the question of whether influenzas antigenic evolution is fundamentally predictable, or if it takes place on a much higher dimensional antigenic space with multiple possible trajectories. We tackle this issue by building a genotype to phenotype map validated on historical hemagglutination inhibition assay data by using machine learning methods. This map uses amino acid physiochemical properties for inference, learning the expected antigenic distance given the differences in polarity and hydrophobicity observed across any two viral sequences, and is thus applicable to newly sampled viruses with previously unseen amino acids. This allows us to accurately blindly predict the antigenic relevance of soon to be vaccine viral strains. We couple the genotype to phenotype map with a molecular evolutionary simulation algorithm to explore the limits of influenzas antigenic evolution and infer to what extent it is in fact predictable. Although we do uncover some canalization of antigenic trajectories, we find that multiple antigenic lineages are equally viable at any one point in time even though typically only one of those trajectories is actually realized.