MIT scientists have developed an artificial intelligence system called VaxSeer that outperforms the World Health Organization’s (WHO) recommended flu vaccine strains in predicting the dominant influenza viruses for upcoming seasons. The WHO’s expert committee meets twice a year to recommend vaccine strains based on global surveillance data, aiming to maximize vaccine effectiveness—the measure of how well the vaccine reduces infection risk in vaccinated versus unvaccinated populations. Each season, agencies like the U.S. Centers for Disease Control and Prevention (CDC) evaluate vaccine effectiveness using observational studies with negative controls. When the selected strain closely matches the circulating virus, inactivated flu vaccines can be 40% to 60% effective. Despite decades of research and monitoring, flu vaccine protection remains suboptimal, prompting MIT researchers to improve strain selection accuracy. VaxSeer uses a deep learning model trained on decades of viral genetic sequences and laboratory test results to simulate how influenza viruses evolve and how vaccines respond. The study, published in Nature Medicine, marks a significant advancement over traditional models that analyze individual amino acid mutations in isolation. Unlike static protein language models that assume fixed viral variants, VaxSeer captures dynamic changes in viral dominance, making it better suited for rapidly evolving viruses like influenza. The system features two core prediction engines: one estimates the likelihood of a virus strain becoming dominant (transmissibility), and the other assesses how well a vaccine can neutralize that strain (antigenicity). Together, they generate a predictive coverage score—a forward-looking metric of how well a given vaccine might perform against future viruses. The score ranges from negative infinity to zero, with values closer to zero indicating better antigenic match and higher predicted effectiveness. In a retrospective 10-year study, researchers compared VaxSeer’s recommendations against WHO’s choices for two major flu subtypes: H3N2 and H1N1. For H3N2, VaxSeer outperformed WHO in 9 out of 10 flu seasons based on the coverage score. For H1N1, it matched or exceeded WHO’s selection in 6 of the 10 seasons. VaxSeer works by first using a protein language model to estimate how quickly different viral strains spread over time. It then models competition among strains to determine which are likely to dominate. The results are fed into a mathematical framework based on ordinary differential equations to simulate viral transmission dynamics. For antigenicity, the system relies on hemagglutination inhibition (HI) assays—standard tests that measure how well antibodies block a virus from binding to human red blood cells, serving as a proxy for antigenic match. “By simulating viral evolution and the interaction between vaccines and viruses, tools like VaxSeer can help public health officials make faster, more accurate decisions—staying ahead in the race between infection and immunity,” said Wenxian Shi, lead author of the study. Currently, VaxSeer focuses only on the hemagglutinin protein of influenza viruses. Future versions could incorporate other proteins like neuraminidase, as well as factors such as population immunity history, manufacturing constraints, and dosage. Expanding the system to other viruses would require large, high-quality datasets tracking viral evolution and immune responses—data that are not always publicly available. However, the team is developing methods to predict viral evolution in low-data environments, leveraging viral family relationships. “Given how fast viruses evolve, current vaccine development often lags behind. VaxSeer is our effort to close that gap,” said Regina Barzilay, MIT professor of engineering and AI and chief researcher at CSAIL. “The work on low-data viral evolution prediction goes beyond flu—it opens new pathways for anticipating disease evolution and designing interventions before escape mutations become a crisis.” Jon Stokes, assistant professor in biochemistry and biomedical sciences at McMaster University in Hamilton, Ontario, praised the study’s broader implications. “This research is not just about better flu vaccines—it’s about a new way of thinking about how we respond to evolving pathogens. It gives us the tools to act proactively, not reactively, in the face of emerging threats.”