Using multiple linear regression and physicochemical changes of amino acid mutations to predict antigenic variants of influenza A/H3N2 viruses

Abstract

Among human influenza viruses, strain A/H3N2 accounts for over a quarter of a million deaths annually. Antigenic variants of these viruses often render current vaccinations ineffective and lead to repeated infections. In this study, a computational model was developed to predict antigenic variants of the A/H3N2 strain. First, 18 critical antigenic amino acids in the hemagglutinin (HA) protein were recognized using a scoring method combining phi (݊) coefficient and information entropy. Next, a prediction model was developed by integrating multiple linear regression method with eight types of physicochemical changes in critical amino acid positions. When compared to other three known models, our prediction model achieved the best performance not only on the training dataset but also on the commonly-used testing dataset composed of 31878 antigenic relationships of the H3N2 influenza virus.

Convolutional Neural Network Based Approach to In Silico Non-Anticipating Prediction of Antigenic Distance for Influenza Virus

“…Cui et al [35] utilized entropy of a specific site, as well as the relationship between mutation occurrence in that site and antigenic variation to recognize the critical position in the sequence. They clustered the well-known AAindex database [36], taking into account the mutual information between physicochemical changes at a critical position and antigenic relationship.…”

“…In modeling of antigenic variants (see, for example, [31,35]), encoding techniques have a direct impact on the predicted results, since they determine how the mutations are represented in the model. To reflect amino acid substitutions in a more descriptive way, we applied physicochemical properties to measure up a quantitative value of the observed mutations.…”

“…Unlike most papers which adopted manual feature engineering [35,62], for example, those based on prior knowledge about antigenic sites and receptor-binding, our approach relies on the advantage of the convolutional neural network framework to provide fully automatic feature extraction by automatically assigning the most relevant features for prediction of the antigenic distance along with the model training.…”

In Silico Vaccine Strain Prediction for Human Influenza Viruses

“…Methods that link antigenic and genetic variation rely primarily on statistical techniques or on information theory [50][51][52][53][54][55][56]. Suzuki [50] estimated antigenic distances between strains of H3N2 based on a model including physicochemical differences between amino acids, the distance between the site and receptor binding site, or to N-linked glycosylation sites, as well as solvent accessibility.…”

“…Cui et al [51] inferred antigenic distances between H3N2 strains and antigenic variants using multivariate regression on multiple physicochemical properties of informative amino acid positions. Ren et al [52] used multivariate regression and feature selection techniques for the HA of H1N1 viruses circulating until 2008 to identify combinations of protein sites that predict antigenic distances between strains.…”