Skip to main content
ukiyo journal - 日本と世界をつなぐ新しいニュースメディア Logo
  • All Articles
  • 🗒️ Register
  • 🔑 Login
    • 日本語
    • 中文
    • Español
    • Français
    • 한국어
    • Deutsch
    • ภาษาไทย
    • हिंदी
Cookie Usage

We use cookies to improve our services and optimize user experience. Privacy Policy and Cookie Policy for more information.

Cookie Settings

You can configure detailed settings for cookie usage.

Essential Cookies

Cookies necessary for basic site functionality. These cannot be disabled.

Analytics Cookies

Cookies used to analyze site usage and improve our services.

Marketing Cookies

Cookies used to display personalized advertisements.

Functional Cookies

Cookies that provide functionality such as user settings and language selection.

AI for Vaccine "Intuition" ─ Will AI Transform Influenza Predictions? MIT's VaxSeer Surpassing WHO in the Future

AI for Vaccine "Intuition" ─ Will AI Transform Influenza Predictions? MIT's VaxSeer Surpassing WHO in the Future

2025年08月30日 08:25

1. Background: The Annual "Looking Ahead" Game

Seasonal influenza vaccines have their strains selected more than half a year in advance. Experts, based on genetic sequences and antigenic data from samples worldwide and using the WHO's GISRS network, estimate which strains will dominate the next season. If the prediction is accurate, the vaccine's effectiveness soars, but if not, people may feel "it doesn't work well despite being vaccinated." The challenge lies in the influenza virus's constant, subtle evolution and its interactions with human immunity and the environment, which can rapidly alter the landscape.


2. What is VaxSeer: Two Prediction Engines + One Score

MIT's VaxSeer extends this "looking ahead" using machine learning. Its core consists of two main components. First, it learns how each mutation contributes to spreadability (dominance) as a "combination" rather than in isolation, using a large-scale protein language model. Second, it estimates how much a vaccine candidate can neutralize future epidemic strains (antigenicity) based on HI (hemagglutination inhibition) test data. These are layered over the dynamics of spread based on ordinary differential equations, ultimately quantifying the vaccine's "fitness" against the expected virus population in future seasons as a "coverage score."


3. How Accurate Was It: A Retrospective 10-Year Review

The research team compared the strains actually chosen by WHO with those recommended by VaxSeer using data from the past 10 years. The results showed that VaxSeer outperformed or equaled WHO in 9 out of 10 seasons for A/H3N2 and 6 out of 10 seasons for A/H1N1. Notably, it identified the candidate strain for the 2016 season a year early, coinciding with WHO's adoption the following year. Furthermore, VaxSeer's coverage score is said to align well with real-world effectiveness estimates published by the CDC, Canada's SPSN, and Europe's I-MOVE. Of course, this is a "retrospective validation" and does not directly indicate the causal effect (true effectiveness) if administered as the vaccine for that year. Nonetheless, for the field where "missed years" significantly increase healthcare burdens, such auxiliary foresight is highly valuable.


4. The Core Mechanism: The Intersection of Evolution and Language

What makes VaxSeer unique is its approach to viewing sequence mutations not as isolated events but as "interactions (epistasis)." Real-world virus evolution cannot be explained by the mere addition of single mutations. Some mutations only cause dominance when combined with others. The language model learns the probabilistic structure of these "combinations" from a vast amount of past sequences. By overlaying a mathematical model representing inter-lineage competition, it gauges the relative advantage of which lineage will become the "next main player." On the antigenicity side, it uses HI test experimental values as a guide to estimate the "distance" between candidate and expected epidemic strains, assessing how much the vaccine can elicit a neutralizing response.


5. Reactions on Social Media: Expectations, Caution, and Operational Realities

Immediately after the announcement, the research community and healthcare-related accounts on social media reacted en masse. MIT CSAIL's official X account introduced it as "learning from decades of data to identify more protective candidates," receiving many positive replies. The first author, Wenxian Shi, also announced the publication of the paper, receiving congratulations from the academic community. On the other hand, accounts following medical AI questioned, "Can AI surpass the judgment of WHO experts?" and highlighted cautious opinions pointing out the evaluation metrics as "surrogate measures" and the lack of prospective clinical validation. Overall, the mainstream perception is that it enhances decision-making transparency as an "aid" or "third opinion" rather than a "replacement." From the clinical and public health side, there are calls for practical application design considering realistic constraints like "manufacturing lead time," "mutations from egg-based cultivation," "supply chain," and "regional data differences."


6. Impact on the Field: How to Use It Effectively

In the short term, VaxSeer's scores could be referenced as a "second opinion" before WHO meetings or national authority selections to detect the risk of a miss early. If signs of a miss are detected, options for policy and operations, such as making supply plans more flexible, adjusting the ratio of alternative manufacturing methods (cellular or recombinant), and strengthening messaging to high-risk groups, can be considered in advance. In the medium term, the key is to expand into a multi-factor model that estimates integrated "effective protection" by incorporating factors other than antigenicity—such as prior immunity, dosage, adjuvants, and the mobility of regional clusters.


7. Limitations and Challenges: AI is Not Omnipotent

First, the evaluation is primarily retrospective. The true test is how much it contributes to reducing actual illness, hospitalization, and mortality when integrated into policy decisions in a prospective environment. Second, biases in the learning data—such as regional, age, and sample collection biases—may distort the generalization of predictions. Third, constraints in the manufacturing process (adaptive mutations in egg cultivation, delays in scaling up) can impair compatibility "from the chosen strain to the final product." Fourth, decision-making is not just about science; it involves areas where transparency, accountability, and social acceptance are questioned. It is essential to disclose the model's assumptions, uncertainties, and sensitivities and position them consistently with expert judgment in governance design.


8. Reasons to Move Forward Nonetheless

Influenza is an "endless enemy." That's why efforts to reduce the probability of a miss by even 1% using data and models are meaningful. VaxSeer connects the principles of evolution with the expressive power of language models, bringing the selection process, which previously relied on human intuition and dispersed indicators, a step closer to a quantitative and reproducible process. Complementarily combining WHO's skilled expert network with AI's foresight capability could be the realistic solution to lighten the next season, even slightly.


Reference Article

MIT's VaxSeer Uses AI to Predict Flu Strains, Beating WHO Picks in Most Seasons Reviewed
Source: https://iafrica.com/mits-vaxseer-uses-ai-to-predict-flu-strains-beating-who-picks-in-most-seasons-reviewed/

Powered by Froala Editor

← Back to Article List

Contact |  Terms of Service |  Privacy Policy |  Cookie Policy |  Cookie Settings

© Copyright ukiyo journal - 日本と世界をつなぐ新しいニュースメディア All rights reserved.