Estimating components of cross-immunity between influenza H3N2 strains based on hemagglutinin sequence data -

Abstract

Global influenza epidemics cause thousands of deaths annually. Vaccination campaigns are an important tool to mitigate effects of influenza epidemics. Because of the fast evolution of the influenza virus, selecting a vaccine strain that confers protection against the main circulating strain remains a key challenge. The WHO uses hemagglutination inhibition (HI) assays, where anti-sera against one strain (serum strain) is set against another strain (virus strain), to determine which strain from a set of vaccine-candidate strains confers best protection against the dominant circulating strains. Even though several studies have shown that cross-immunity can be predicted from hemagglutinin (HA) sequences, current techniques still use HI assays. One shortcoming of sequence-based predictions is that they rely only on the differences between two HA sequences, and thus assume symmetry in cross immunity. Our study introduces a method for sequence-based prediction of cross-immunity that relaxes the symmetry assumption. In our method, each amino acid of the virus strain HA, each amino acid of the serum strain HA and each amino acid difference between the virus and serum strain HA were included as a potential predictor variable and log-transformed HI titers were used as response variable. Regression coefficients were estimated via elastic net regression with cross-validation. The data was split in a training and validation set. The training set was used to estimate regression coefficients and the validation set was used to predict HI titers based on estimated coefficients and compare them to the actual HI titer values. The coefficients for the correlations between estimated and actual HI titers were 0.72 and 0.67 for training and validation sets, respectively. Most amino acid positions that received non-zero regression coefficients fell within the epitope regions or were in close proximity to those regions on the 3D structure of the HA protein. Our results suggest that the proposed model can predict HI titers and find antige