The challenges of vaccine strain selection

New measures of influenza virus fitness could improve vaccine strain selection through more accurate forecasts of the evolution of the virus.

S cientists have known since the 1940s that influenza vaccines that perform well one year can be rendered ineffective after the influenza virus mutates. However, despite decades of investment in global surveillance, pathogen sequencing technologies and basic research (Figure 1), vaccines for seasonal influenza have the lowest and most variable performance of any vaccine licensed for use in the United States (CDC, 2016). Now, in eLife, John Huddleston of the Fred Hutchinson Cancer Research Center (FHCRC) and the University of Washington, Trevor Bedford of the FHCRC, and colleagues in the United States, United Kingdom, Japan, Australia and Switzerland present an open-source framework that synthesizes a decade's worth of innovations in bioinformatics and technology to advance data-driven vaccine design (Huddleston et al., 2020).
Influenza A and influenza B viruses cause seasonal epidemics every winter. Seasonal influenza A viruses include two different subtypes, H1N1 and H3N2, where H and N (short for hemagglutinin and neuraminidase) are proteins found on the surface of the virus. The human immune system protects the body against influenza infection by producing antibodies that can recognize these proteins. However, the influenza virus mutates frequently, including at sites that affect the immune system's ability to detect the virus. This process -called 'antigenic drift' -helps the virus infect new hosts and spread in populations that previously had immunity to influenza. Indeed, antigenic drift can lead to new strains of the virus that completely displace the currently circulating strains in a matter of months.
To keep pace with antigenic drift, the composition of influenza vaccines must be updated continually. Influenza vaccines contain three or four components that protect against various strains representing the different subtypes. Scientists convene twice a year at the World Health Organization (WHO) to predict which strains will have the highest fitness and therefore dominate the next year's flu season. H3N2 viruses evolve particularly fast and unpredictably compared to other seasonal flu viruses. Because the composition of the vaccine has to be decided a year in advance to allow doses to be manufactured, H3N2 vaccine strains have failed to match naturally circulating strains in six of the past fifteen flu seasons ( Figure 1).
For decades, vaccine strain selection has been primarily informed by data from 1950s-era serological assays, which provide a phenotypic Copyright Perofsky and Nelson. This is an open-access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication. measure of how immune systems exposed to recently circulating viruses would see a novel strain. However, the assays have certain disadvantages -they are labor intensive, inconsistent across labs, not publicly available, and difficult to interpret or scale up. This means that these phenotypic measures are only available for a small subset of viruses. To remedy this issue, Huddleston et al. use a phylogenetic model (which includes available serological data and sequence data as inputs) to make predictions for the thousands of strains for which serological information is not available (Bedford et al., 2014;Neher et al., 2016;Smith et al., 2004).
Huddleston et al. compare how antigenic phenotypes from serological assays perform against five newer measures of virus fitness in forecasting future H3N2 virus populations, and find that two of their models provide better forecasts than WHO vaccine strain selections. Moreover, they have now integrated their forecasts for H3N2 into nextstrain.org, an opensource platform that scientists and policymakers use to track the real-time evolution of a wide range of pathogens (Hadfield et al., 2018;Neher and Bedford, 2015). Nextstrain provides a platform to make influenza vaccine strain selection more data-driven, systematic and transparent, and to allow new forecasting methods to be integrated as they show promise.
How does one predict the fitness of an influenza virus? Most mutations are harmful for influenza viruses, except for a subset of beneficial mutations that lead to antigenic drift. For decades researchers have relied on a list of sites in the genome where seemingly beneficial mutations occur to measure antigenic drift and viral fitness (Bedford et al., 2014;Bush et al., 1999;Shih et al., 2007). However, Huddleston et al. find that serological assays (Neher et al., 2016) continue to be more useful than sequence-onlybased measures when making forecasts of future virus populations. Measures of viral fitness based on genetic sequences could not accurately predict H3N2 evolution in recent years due to the emergence of multiple co-circulating strains and the sudden decline of a dominant strain in 2019. While no method predicts the right vaccine strain every time, serology-based methods appear to outperform other approaches.
Over time, alternative approaches to measuring virus fitness will continue to be refined and  (Knossow et al., 1984;Wiley and Skehel, 1987) and predicting the evolution of H3 (Bush et al., 1999)  may become integrated into vaccine strain selection. For example, Huddleston et al. could not include a new serological assay based on virus neutralization in their framework as data from this assay were only available over a short period of time, but it could be integrated as data accrue. Other incremental improvements could be beneficial when used in combination with serological data. For example, how fast a strain is spreading globally can be measured from branching patterns in the phylogenetic tree (Neher et al., 2014). 'Mutational load' (that is, the total number of mutations in sites unrelated to immune detection) provides a simple inverse measure of viral fitness (Luksza and Lässig, 2014), while a technique called deep mutational scanning measures whether experimentally induced mutations have beneficial or harmful effects (Lee et al., 2018), However, as with other sequence-based approaches, the fact that mutations have different effects in different genetic backgrounds may be a disadvantage.
Going forward, the COVID-19 pandemic could disrupt the ecology of flu viruses in the years ahead, and it will be interesting to observe how predictive models fare in a highly perturbed system with no historical precedent. SARS-CoV-2 viruses may also experience post-pandemic strain turnover that requires periodic updates to any COVID-19 vaccine, and it should be possible to adapt platforms built for influenza forecasting to make forecasts for SARS-CoV-2 and other pathogens.

Note
Disclaimer: The conclusions of this study do not necessarily represent the views of the NIH or the US government. Competing interests: The authors declare that no competing interests exist.