Choosing the right influenza strains months before flu season remains a high-stakes guess. A new AI system from MIT aims to change that. Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory and the Abdul Latif Jameel Clinic for Machine Learning in Health unveiled VaxSeer, a model that forecasts dominant flu strains and recommends vaccine candidates well ahead of time. An open-access study in Nature Medicine published today details the approach.
VaxSeer uses deep learning trained on decades of viral genetic sequences and lab results to simulate how influenza evolves and how candidate vaccines are likely to perform. Unlike traditional models that assess one mutation at a time, VaxSeer’s large protein language model learns how combinations of mutations influence viral “dominance” and vaccine match over time.
The system has two engines: one estimates how likely a strain is to spread, the other estimates how well a vaccine will neutralize that strain (antigenicity). Together they produce a predicted coverage score, a forward-looking measure of how closely a vaccine will match future circulating viruses. Scores closer to zero indicate a better match.
In a 10-year retrospective analysis, VaxSeer’s recommendations for A/H3N2 outperformed the World Health Organization’s selections in 9 of 10 seasons based on empirical coverage scores derived from observed dominance and hemagglutination inhibition tests. For A/H1N1, VaxSeer matched or exceeded WHO choices in 6 of 10 seasons, including identifying a strain for 2016 that WHO selected the following year. The model’s predictions correlated with real-world vaccine effectiveness reported by the CDC, Canada’s Sentinel Practitioner Surveillance Network, and Europe’s I-MOVE program.
Under the hood, VaxSeer models strain competition and spread with ordinary differential equations and estimates antigenic performance from lab assay data. The current version focuses on the HA protein; future work may incorporate NA, immune history, manufacturing constraints, and dosage. The team is also developing methods for low-data settings to extend predictive evolution beyond influenza.
“By modeling how viruses evolve and how vaccines interact with them, AI tools like VaxSeer could help health officials make better, faster decisions,” said lead author Wenxian Shi, an MIT PhD student and CSAIL researcher. Senior author Regina Barzilay said the goal is to help therapeutic development keep pace with rapid viral change.
External experts say the implications reach further. Jon Stokes of McMaster University noted that similar predictive frameworks could anticipate trajectories of antibiotic resistance or drug-resistant cancers, enabling earlier, better-targeted interventions.
The work was supported in part by the U.S. Defense Threat Reduction Agency and the MIT Jameel Clinic.
