Influenza viruses pose a significant global public health threat, and vaccines are the primary pharmaceutical intervention. However, the hemagglutinin (HA) protein, the main target of influenza vaccines, undergoes antigenic changes, requiring annual updates to vaccine composition to maintain efficacy. This process, coordinated by the WHO Global Influenza Surveillance and Response System (GISRS), is resource-intensive and time-consuming. GISRS recommends appropriate vaccine viruses and provides candidate vaccine viruses (CVVs) to manufacturers. An ideal CVV must possess the correct antigenic properties, maintain these properties throughout production, and exhibit high growth yield. Delays in identifying high-yielding CVVs can lead to vaccine mismatches and supply shortages, as exemplified by the 2003-2004 and A(H1N1)pdm09 pandemic seasons. Current strategies for improving CVV yield, such as additional passages in eggs or cells, and reassortment with high-yield donor strains, are time-consuming (up to 6 months) and may introduce unwanted antigenic changes. Therefore, a rapid method for identifying high-yield influenza vaccine viruses directly from clinical samples is highly desirable. While several computational models exist for identifying antigenic variants using genomic sequences, none can directly predict both antigenic match and high yield from genetic sequences. This study addresses this limitation by introducing MAIVESS.

Existing computational models for influenza antigenic variant identification primarily focus on using genomic sequences to predict antigenic properties. These models, however, lack the ability to predict both antigenicity and high yield directly from genetic sequences. The authors cite several previous studies that have developed computational models to predict influenza antigenic variants, but highlight the limitations of these models in directly identifying antigenically matched and high-yield viruses based on genetic sequences. The review sets the stage for the introduction of MAIVESS as a novel approach to overcome these challenges.

The researchers developed MAIVESS, a machine-learning assisted influenza vaccine strain selection framework. MAIVESS uses machine learning algorithms to predict three key influenza virus properties: antigenicity, growth yield, and receptor binding. Multiple machine learning models were compared, with MTL-GGSL (multi-task learning group-guided sparse learning) selected for predicting antigenicity and glycan binding, and GHSM (generalized hierarchical sparse model) for predicting growth yield. To train and validate the models, a library of A(H1N1)pdm09 mutant viruses was generated targeting the HA receptor-binding site (RBS). These mutants were subjected to antigenic analysis (hemagglutination inhibition [HAI] assays), yield analysis (in MDCK cells and embryonated chicken eggs), and receptor-binding profiling (glycan microarrays). The phenotypic data were then used to train and test MAIVESS. The framework uses the learned features to score candidate vaccine viruses (CVVs) based on antigenic properties and yield in eggs and/or cells. The MAIVESS framework is accessible through both GitHub and a webserver. The study then applied MAIVESS to a dataset of A(H1N1)pdm09 viruses to identify potential CVVs, which were subsequently validated experimentally. Biolayer interferometry (BLI) was used to confirm glycan binding profiles. Structural modeling was employed to investigate the effect of specific amino acid substitutions on glycan binding.

MAIVESS effectively predicted both antigenicity and yield phenotypes using HA protein sequences. The analysis of 189 A(H1N1)pdm09 mutants revealed that substitutions near the HA RBS, while generally not altering antigenicity, could significantly impact yield in both cells and eggs. The highest-yield mutants showed a >100-fold increase compared to the wild type. Analysis identified 30 residues associated with antigenicity and 38 residues associated with yield. Glycan binding profiling showed that high-yield mutants often displayed a broader binding specificity to sialylated glycans, particularly 3'SLN, in addition to the commonly bound 6'SLN. Application of MAIVESS to a large dataset of A(H1N1)pdm09 sequences (2009-2020) identified numerous high-yield candidates, many of which were antigenically matched to circulating strains. Notably, the proportion of high-yield strains increased significantly after 2018. Experimental validation of four MAIVESS-selected CVVs confirmed their high yield in both eggs and cells (at least 100-fold higher than the wild type) and predicted antigenicity.

MAIVESS successfully addresses the limitations of existing methods by directly predicting both antigenicity and yield from HA sequences, enabling rapid selection of optimal CVVs. The finding that increased glycan binding diversity, particularly to 3'SLN, correlates with high yield suggests a potential mechanism for the acquisition of this trait in naturally circulating viruses. The model's ability to identify high-yield strains that are antigenically matched to circulating strains is particularly significant for vaccine development. The increased prevalence of high-yield strains after 2018 warrants further investigation. While the model currently focuses on HA, future iterations could incorporate NA sequences and human serological data for improved accuracy. The use of ferret antisera for antigenicity testing is a limitation, as human immune responses can differ.

MAIVESS offers a significant advancement in influenza vaccine strain selection, reducing the time required for identifying optimal CVVs from months to days. Its ability to predict both antigenicity and yield directly from clinical samples accelerates vaccine production and ensures timely vaccine supply. Future research should focus on incorporating NA sequence data and human serological data into the model, further enhancing its predictive capabilities.

The study primarily focused on A(H1N1)pdm09 viruses. The generalizability to other influenza subtypes needs further investigation. The use of ferret antisera for antigenicity assessment might not fully capture the nuances of human immune responses. Additionally, the study did not fully investigate all factors influencing viral yield, such as innate immune responses and overall viral fitness.