Every year, global health experts face high stakes decisions. Which flu strains should be in the vaccine next season? The choice must be made a few months before the flu season begins, and often feels like a race against the watch. If the selected strain matches the circulating strain, the vaccine can be very effective. However, if forecasts are off, protection can be significantly reduced, leading to (potentially preventable) illnesses, putting a strain on the health system.
This challenge has become even more familiar to scientists during the Covid-19 pandemic. Just as vaccines are rolling out, remember time (and many times) when new variations emerge. Influenza behaves like this as it is not violent, and is constantly mutating predictively. This makes it difficult to go ahead and, therefore, to design a vaccine that remains protected.
To reduce this uncertainty, scientists at MIT’s Computer Science and Artificial Intelligence Institute (CSAIL) and MIT Abdul Latif Jameel Clinic make vaccine choices more accurate and reliant on guessing for machine learning in health conditions. They created an AI system called Vaxseer, designed to predict dominant flu strains and identify the most protective vaccine candidates a few months ago. The tool uses a deep learning model trained with decades of virus sequencing and lab test results to simulate how influenza viruses evolve and the vaccine response.
Traditional evolutionary models often independently analyse the effects of single amino acid mutations. “Vaxseer employs a large-scale protein language model to learn the relationship between domination and the combined effects of mutations,” explains Wenxian Shi, a doctoral student in the Department of Electrical Engineering and Computer Science at MIT, CSAIL researcher and lead author of a new paper on the work. “Unlike existing protein language models that assume the static distribution of viral mutants, they model dynamic dominant shifts and are suitable for rapidly evolving viruses like influenza.”
An open access report on this study was published today Natural medicine.
The future of influenza
Vaxseer has two core prediction engines. One is to estimate how likely each strain of virus is to spread (dominate), and the other is to estimate how effectively a vaccine neutralizes its strain (antigenicity). Together, they generate a predicted coverage score. It is a measure of future outlook as to how well a particular vaccine will work against future viruses.
The score scale can range from infinite negative to zero. The closer the score is to zero, the better the antigenic match of the vaccine strain against circulating viruses. (You can imagine it as a kind of “distance” negative.)
In a 10-year retrospective study, researchers evaluated Vaxseer’s recommendations for what the World Health Organization (WHO) created for two major influenza subtypes A/H3N2 and A/H1N1. For A/H3N2, Vaxseer’s choice outperformed the WHO in 9 seasons of 10 seasons based on a retrospective empirical coverage score (a proxy index of vaccine efficacy calculated from observed advantages from past seasons and experimental HI test results). The team used this to evaluate vaccine choices. This is because efficacy can be utilized only with vaccines that are actually given to the population.
For A/H1N1, they outperformed and matched WHO in six of the 10 seasons. In one notable case, during the 2016 flu season, Vaxseer identified tensions that WHO did not choose until the following year. Model predictions showed strong correlations with estimates of actual vaccine efficacy, as reported by the CDC, the Canadian Sentinel Practitioner Surveillance Network, and the European i-move program. Vaxseer’s predicted coverage score was closely matched with public health data on flu-related illnesses and medical visits prevented by vaccination.
So, how does Vaxseer understand all this data exactly? Intuitively, this model first uses a protein language model to estimate how quickly viral strains expand over time, and then determine its advantage by taking into account competition between different strains.
When the models calculate insights, they are connected to a mathematical framework based on what is called ordinary differential equations, simulating virus spreads over time. As for antigenicity, the system estimates how well a particular vaccine strain performs in a common lab test called hemagglutination inhibition assay. This measures how effectively the antibodies are inhibited from inhibiting the binding of the virus to human erythrocytes. This is a widely used proxy for antigen matching/antigenicity.
surpass evolution
“By modeling how viruses evolve and how vaccines interact, AI tools like Vaxseer can help health officials make better, faster decisions and stay one step ahead in the race between infection and immunity,” Shi says.
Vaxseer currently focuses solely on the HA (hemagglutinin) protein of the influenza virus, the major antigen of influenza. Future versions can incorporate other proteins such as NA (neuraminidase), as well as factors such as immune history, production constraints, and dosage levels. Applying the system to other viruses also requires a large, high-quality dataset that tracks both the evolution of the virus and the immune response. This is data that is not always public. However, the team is currently working on ways to predict the evolution of low-data regimes, which are built on relationships between viral families.
“Given the speed of virus evolution, current therapeutic development is often delayed. Vaxseer is our attempt to catch up,” says Regina Barzilay, MIT’s Faculty of AI and Health, AI Lead at Jameel Clinic, and Chief Investigator at CSAIL.
“This paper is impressive, but perhaps what excites me even more,” said John Stokes, assistant professor at the Department of Biochemistry and Biomedical Sciences at McMaster University in Hamilton, Ontario. “Imagine that the impact is far beyond influenza. Imagine how antibiotic-resistant bacteria or drug-resistant cancers can be predicted. Both can adapt quickly. This type of predictive modeling opens up a powerful new way of thinking about how disease changes, gives the opportunity to take one step, and design clinical interventions.”
Shi and Barzilay have written papers for MIT Csail Postdoc Jeremy Wohlwend ’16, Meng ’17, PhD ’25, and recent Csail affiliates Menghua Wu ’19, Meng ’20, and PhD ’25. Their work was supported in part by the US Defense Threat Reduction Agency and the MIT Jameel Clinic.
