Each year, global health experts deal with a high rate decision: which flu strains should go to the next seasonal vaccine? The choice must be made a month in advance, long before the start of the flu season and can often feel like a race with a clock. If selected strains match those that circulate, the vaccine will probably be very effective. But if the anticipation is turned off, the protection can drop significantly, which leads to (potentially possible to prevent) disease and burden of healthcare systems.
This challenge became even more known to scientists in the years during the Covid-19 pandemic. Return to the time (and again and again), when new variants appeared when vaccines were implemented. The flu behaves like a similar, noisy cousin, constantly and unpredictable. This makes it difficult to stay at the forefront, and therefore it is more difficult to design vaccines that remain protective.
To reduce this uncertainty, scientists at Mit's Computer Science and Artificial Intelligence Laboratory (CSAIL) and the myth of Abdul Latif Jameel Clinic for Machine Machine learning in health to make vaccine selection more accurate and less dependent on guessing. They created the AI system called Vaxseer, designed to predict the dominant influenza strains and identify the most protective vaccine candidates, several months in advance. The tool uses deep learning models trained in decades of viral sequences and laboratory test results to simulate the evolution of influenza virus and how vaccines will react.
Traditional evolution models often independently analyze the influence of mutations of individual amino acids. “Vaxseer accepts a large model of protein language to learn the relationship between domination and combinatorial effects of mutations,” explains Wenxian Shi, a PhD student at the Department of Electrical Engineering and Computer Science MIT, CSAIL researcher and main author of the new article on work. “In contrast to existing models of protein languages, which assume a static distribution of viral variants, we model dynamic shifts of dominance, making it better suitable for rapidly developing viruses such as flu.”
Some Report on open research available It was published today in Nature Medicine.
The future of the flu
Vaxseer has two basic forecast engines: one that estimates, as it is likely, that each viral strain is spread (domination), and the other, which estimates how effectively the vaccine neutralizes this strain (antigenicity). Together, they produce the expected range of the range: the future, how well a given vaccine will probably work in relation to future viruses.
The scale of the result can be from infinite negative to 0. The closer to the result to 0, the better the antigenic match of strains to the circulating viruses. (You can imagine it as a negative “distance”.)
In a 10-year retrospective study, scientists evaluated Vaxseer recommendations against the recommendations of the World Health Organization (WHO) for two main fluid subtypes: A/H3N2 and A/H1N1. In the case of A/H3N2, Vaxseer choices have achieved better results than nine seasons, based on retrospective empirical coverage results (replacement record of vaccine effectiveness, calculated on the basis of observed domination from previous seasons and results of experimental HI tests). The team used this to assess the selection of vaccines, because effectiveness is only available for vaccines actually transferred to the population.
In the case of A/H1N1, it exceeded or matched WHO for six to 10 seasons. In one significant case, in the 2016 flu season, Vaxseer identified a strain that was not chosen by WHO until the next year. The model forecasts also showed a strong correlation with real estimates of the effectiveness of vaccines, as reported by CDC, the Canadian Sentinel and I-Move program supervision network in Europe. Estimated results of Vaxseer covering strictly in line with public health data on diseases related to flu and visits to doctor prevented by vaccination.
How exactly Vaxseer has the meaning of all this data? Intuitively, the model first estimates how quickly the viral strain spreads over time using a protein language model, and then determines its dominance by taking into account the competition between different strains.
After calculating the model of its observations, they are connected to the mathematical frame based on something that is called ordinary differential equations to simulate the spread of the virus over time. In the case of antigenicity, the system estimates how well a given vaccine strain will perform in a common laboratory test called hemaglutation inhibition test. It measures how antibodies can effectively brake the virus from the bond with human red blood cells, which is widely used proxy for antigenic fit/antigenicity.
Ahead evolution
“Modeling, how viruses evolve and how vaccines with them interact with them, AI tools, such as Vaxseer, can help health officials to make better, faster decisions – and stay a step ahead in the race between infection and immunity,” says Shi.
Vaxseer is currently focused only on the protein of half virus (hemaglutinin), the main influenza antigen. Future versions may include other proteins, such as (neuraminidase) and factors such as immune history, production restrictions or dosage levels. The use of a system for other viruses would also require large, high-quality data sets that follow both viral evolution and immune-dane answers that are not always publicly available. The team, however, is currently working on methods that can predict viral evolution in low -data regimes based on relationships between viral families
“Considering the speed of viral evolution, the current therapeutic development is often lagging behind. Vaxseer is our attempt to catch up,” says Regina Barzilay, an outstanding professor School of Engineering for AI and health in myth, and the leading Jameel clinic and the main CSAIL researcher.
“This article is impressive, but perhaps even more excited about the ongoing work of the team over predicting viral evolution in low dates,” says assistant Professor Jon Stokes from the Faculty of Biochemistry and Biomedic Sciences at the University of McMaster at Hamilton, Ontario. “Implications go far beyond flu. Imagine that you can predict how antibiotics or drug -resistant crayfish bacteria can evolve, both can adapt quickly. This type of predictive modeling becomes a serious problem.”
Shi and Barzilay wrote an article with the MIT CSAIL Postdoc Jeremy Wohlwend '16, Meng '17, PhD '25 and the recent partner of CSail Menghua Wu '19, Meng '20, PhD '25. Their work was partly supported by the American Agency for Reduction of Defense Threats and the myth of Jameel Clinic.
