MIT researchers have developed an AI tool called VaxSeer to improve the accuracy of seasonal flu vaccine strain selection. Each year, global health experts must predict which influenza strains will dominate months before flu season begins, a process that often relies on limited data and can lead to mismatched vaccines. The new AI system aims to reduce uncertainty by forecasting dominant flu strains and identifying the most effective vaccine candidates well in advance. VaxSeer uses deep learning models trained on decades of viral genetic sequences and laboratory test results. Unlike traditional methods that assess individual mutations in isolation, VaxSeer employs a large protein language model to understand how combinations of mutations affect a virus’s ability to spread and evade immunity. It also accounts for dynamic shifts in viral dominance over time, making it better suited for fast-evolving viruses like influenza. The tool has two main components: one predicts how likely a strain is to become widespread (dominance), and the other estimates how well a vaccine can neutralize that strain (antigenicity). Together, they generate a coverage score—a forward-looking metric indicating how well a vaccine is expected to perform. The closer the score is to zero, the better the match between the vaccine and future circulating strains. In a 10-year retrospective analysis, VaxSeer outperformed the World Health Organization’s recommendations in nine out of ten seasons for the A/H3N2 subtype, and matched or exceeded them in six out of ten for A/H1N1. In 2016, the model identified a strain later chosen by the WHO only in the following year. Its predictions also closely aligned with real-world vaccine effectiveness data from the CDC, Canada’s Sentinel Practitioner Surveillance Network, and Europe’s I-MOVE program. VaxSeer works by first estimating viral spread using a protein language model, then modeling strain competition over time through a system of ordinary differential equations. For antigenicity, it predicts how well a vaccine would perform in the hemagglutination inhibition assay—a standard lab test measuring antibody effectiveness in blocking viral entry into human cells. The system currently focuses on the hemagglutinin (HA) protein, the primary target of flu vaccines. Future versions may include neuraminidase (NA) and incorporate factors like prior immunity, manufacturing limits, and dosage. Expanding the tool to other viruses would require large, high-quality datasets tracking both viral evolution and immune responses—data that are often scarce. The research team, led by MIT CSAIL and the Abdul Latif Jameel Clinic for Machine Learning in Health, includes Wenxian Shi, Regina Barzilay, Jeremy Wohlwend, and Menghua Wu. Their work was supported by the U.S. Defense Threat Reduction Agency and MIT Jameel Clinic. Experts say VaxSeer represents a major step forward in using AI to stay ahead of viral evolution. As Barzilay noted, “This paper is impressive, but what excites me even more is the team’s ongoing work on predicting viral evolution in low-data settings.” The implications could extend beyond flu to antibiotic-resistant bacteria and drug-resistant cancers—diseases that also evolve rapidly. By anticipating these changes, such tools could help design treatments before resistance becomes widespread.