Colourised transmission electron micrograph showing H1N1 influenza virus particles. Image: NIAID/Flickr, CC BY 2.0.
Scientists 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, vaccines for seasonal influenza have the lowest and most variable performance of any vaccine licensed for use in the United States.
Now, in eLife, John Huddleston of the Fred Hutchinson Cancer Research Centre (FHCRC) and the University of Washington, Trevor Bedford of the FHCRC, and colleagues in the US, UK, Japan, Australia and Switzerland present an open-source framework that synthesises a decade’s worth of innovations in bioinformatics and technology to advance data-driven vaccine design.
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 recognise 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 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 15 flu seasons.
For decades, vaccine strain selection has been primarily informed by data from 1950s-era serological assays, which provide a phenotypic measure of how immune systems exposed to recently circulating viruses would see a novel strain. However, the assays have certain disadvantages – they are labour-intensive, inconsistent across labs, not publicly available and are difficult to interpret or scale up.
(Editor’s note: “The term ‘phenotype’ refers to the observable physical properties of an organism” {source}.)
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.
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, an open-source platform that scientists and policymakers use to track the real-time evolution of a wide range of pathogens. 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.
Also read: How a Coronavirus Mutation in Minks Could Disrupt Vaccine Development
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. However, Huddleston et al. find that serological assays continue to be more useful than sequence-only based 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 may become integrated into vaccine strain selection. For example, Huddleston et al. could not include a new serological assay based on virus neutralisation 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. Mutational load (that is, the total number of mutations in sites unrelated to immune detection) provides a simple inverse measure of viral fitness, while a technique called deep mutational scanning measures whether experimentally induced mutations have beneficial or harmful effects, 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.
This article was originally published on eLife and has been republished here under a Creative Commons Zero license. The reference in the text have been omitted for ease of reading; they are available to view in the original.
Amanda C. Perofsky and Martha I. Nelson are with the Division of International Epidemiology and Population Studies, Fogarty International Centre, National Institutes of Health, US.