Every year, global health experts face a life-or-death decision: Which flu strains should be included in next season’s vaccine? This decision must be made months in advance, before the season has even begun. If chosen correctly, the vaccine will be highly effective. But if it goes wrong, protection will be significantly reduced, leading to a flood of preventable cases and putting enormous pressure on health systems.

Professor Regina Barzilay (left) and graduate student Wenxian Shi. Photo: MIT News

This challenge has become all the more familiar during the Covid-19 pandemic, where new variants have emerged just as vaccines are being rolled out. Influenza behaves similarly – like a “noisy sibling,” constantly and unpredictably mutating, leaving vaccine design one step behind.

To reduce uncertainty, scientists at the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT Abdul Latif Jameel Clinic for Machine Learning in Healthcare created an AI system called VaxSeer. The tool predicts the dominant future flu strain and identifies the best vaccine candidates to protect months before an outbreak. VaxSeer was trained on decades of data including virus genetic sequences and lab test results to simulate how the virus evolves and responds to vaccines.

Unlike traditional evolutionary models that analyze individual amino acid mutations, VaxSeer uses a “protein language model” to learn the relationship between dominance and the combined effects of multiple mutations. “We simulate the dynamic change of dominance, which is more appropriate for rapidly evolving viruses like influenza,” said Wenxian Shi, a PhD student at MIT and lead author of the study.

How does VaxSeer work?

This tool has two main prediction engines:

Dominance: An estimate of the likelihood that a strain of influenza will spread.
Antigenicity: Predicts how effective the vaccine is at neutralizing that strain.
Combining the two factors, VaxSeer generates a “predictive coverage score,” which shows how closely the vaccine matches future strains of the virus. The closer this score is to zero, the better the match.

In a 10-year retrospective study, the MIT team compared VaxSeer's recommendations with the World Health Organization's (WHO) choices for two major influenza subtypes: A/H3N2 and A/H1N1.

For A/H3N2, VaxSeer's recommendation outperformed WHO's in 9/10 epidemic seasons.
For A/H1N1, the system was equal to or better than WHO in 6/10 seasons.
Notably, in the 2016 flu season, VaxSeer picked out a strain that the WHO wouldn't include in a vaccine until the following year.

VaxSeer's predictions also correlate closely with real-world vaccine efficacy data from the CDC (USA), the Practice Surveillance Network in Canada, and the I-MOVE program in Europe.

Racing with virus evolution

VaxSeer estimates the spread rate of each virus strain using a protein language model, then calculates dominance based on competition between strains. Next, the data is fed into a mathematical framework based on differential equations to simulate the spread.

Photo of article 78.jpg

For antigenicity, VaxSeer predicts vaccine efficacy through the hemagglutination inhibition test (HI test), a common measure of antigenicity.

“By modeling viral evolution and vaccine responses, AI tools like VaxSeer can help health officials make faster and better decisions, staying one step ahead in the race between infection and immunity,” Shi asserted.

VaxSeer currently focuses on the HA (hemagglutinin) protein, the main influenza antigen. Future versions could include the NA (neuraminidase) protein, immune history, manufacturing processes, or dosage. The team is also developing a method to predict virus evolution in the absence of data, based on relationships between virus families.

“VaxSeer is our attempt to keep up with the rapid pace of virus evolution,” said Regina Barzilay, MIT Distinguished Professor of AI and Medicine and co-author of the study.

Jon Stokes, assistant professor at McMaster University (Canada), commented: “The amazing point is not only the current results, but also the potential to extend to other areas: predicting the evolution of drug-resistant bacteria or treatment-resistant cancers. This is a completely new approach, allowing medical solutions to be designed before the disease has a chance to overcome the barrier.”

(According to MIT)

Source: https://vietnamnet.vn/mit-phat-trien-cong-cu-ai-du-doan-virus-cum-cuu-hang-trieu-ca-benh-2439275.html