MIT Develops Artificial Intelligence for More Accurate Flu Vaccines

Every year, health experts face a challenging task: selecting the right influenza virus strains for the seasonal vaccine. This selection must take place months before the start of flu season, often making it a race against time. If the selected strains match those actually circulating, the vaccine is highly effective. But if the predictions miss the mark, protection decreases and the burden on healthcare systems increases. This problem became even more apparent during the COVID-19 pandemic, when new virus variants emerged months after vaccines had been deployed. Influenza behaves similarly—it constantly mutates and is unpredictable, making it difficult to design vaccines that remain protective.

Scientists from the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT Abdul Latif Jameel Clinic for Machine Learning in Health set out to reduce this uncertainty. They developed an artificial intelligence system called VaxSeer, which predicts dominant flu strains and identifies the most protective vaccine candidates months in advance. The tool uses deep learning models trained on decades of virus sequences and laboratory test results to simulate how the influenza virus may evolve and how vaccines will respond to it.

How Does VaxSeer Work?

Traditional evolutionary models often analyze the impact of individual amino acid mutations separately. VaxSeer, however, uses a large protein language model to learn the relationships between dominance and the combined effects of mutations. Unlike existing protein language models, which assume a static distribution of virus variants, VaxSeer models dynamic shifts in dominance, making it better suited to rapidly evolving viruses such as influenza. According to Wenxian Shi, an MIT doctoral student in electrical engineering and computer science and the study's lead author, this approach enables better adaptation to rapidly changing viruses.

VaxSeer has two main prediction engines: one estimates how likely each virus strain is to spread (dominance), and the other estimates how effectively a vaccine will neutralize that strain (antigenicity, or the ability to induce an immune response). Together, they produce a predicted coverage score that measures how well a given vaccine is likely to perform against future viruses. The score ranges from negative infinity to zero—the closer it is to zero, the better the vaccine's antigenic match with circulating viruses.

The model first estimates a strain's rate of spread using a protein language model and then determines its dominance while accounting for competition among different strains. These insights are incorporated into a mathematical framework based on ordinary differential equations that simulate the virus's spread over time. For antigenicity, the system estimates how well a vaccine performs in a standard laboratory test called the hemagglutination inhibition assay, which measures how effectively antibodies prevent the virus from binding to human red blood cells.

Results and Comparison With Real-World Data

In a decade-long retrospective study, the researchers evaluated VaxSeer's recommendations against those of the World Health Organization (WHO) for two major influenza subtypes: A/H3N2 and A/H1N1. For A/H3N2, VaxSeer's selections outperformed the WHO's in nine out of ten seasons based on retrospective empirical coverage scores, which serve as a proxy metric for vaccine efficacy. For A/H1N1, it outperformed or matched the WHO in six out of ten seasons. In one instance, for the 2016 season, VaxSeer identified a strain that the WHO did not select until the following year.

The model's predictions also showed a strong correlation with real-world estimates of vaccine effectiveness, as reported by the CDC (Centers for Disease Control and Prevention), Canada's Sentinel Practitioner Surveillance Network, and the European I-MOVE program. VaxSeer's predicted coverage scores closely matched data on the prevention of influenza-related illnesses and medical visits through vaccination.

According to Regina Barzilay, a professor at the MIT School of Engineering and a principal investigator at CSAIL, VaxSeer helps health authorities make better and faster decisions to stay one step ahead of infections and immunity. The study was published in the journal Nature Medicine and was supported in part by the U.S. Defense Threat Reduction Agency.

The Future of VaxSeer

VaxSeer currently focuses only on the HA (hemagglutinin) protein, the primary influenza antigen. Future versions could include other proteins such as NA (neuraminidase), as well as factors such as the population's immune history, manufacturing constraints, or dosage levels. Applying the system to other viruses would require large, high-quality datasets tracking viral evolution and immune responses, which are not always publicly available. The team is now working on methods for predicting viral evolution in low-data settings based on relationships among virus families.

According to Jon Stokes, an assistant professor in the Department of Biochemistry and Biomedical Sciences at McMaster University in Hamilton, Ontario, this work is impressive and paves the way for predicting evolution in other challenges, such as antibiotic-resistant bacteria or drug-resistant cancers. VaxSeer is designed as a screening tool for prioritizing candidates for further laboratory validation, not as a replacement for the WHO's current process. Its predictions could reveal more effective strains that have not yet been tested and contribute to greater vaccine effectiveness, potentially reducing illness and saving lives.