Predicting the antigenic evolution of seasonal influenza viruses using phylogenetic convergence

This study demonstrates that analyzing convergent evolution in large phylogenies can predict major antigenic shifts in H3N2 influenza viruses over a year in advance by identifying specific receptor-binding site substitutions under positive selection before they become dominant.

The Gist

Imagine the flu virus as a master thief trying to break into a bank vault. The vault is your immune system, and the lock is a specific protein on the virus's surface called Hemagglutinin (HA). Every year, we build a new "security system" (the flu vaccine) based on what the thief looked like last year. But the thief is clever; they change their disguise (mutate) to slip past the guards.

The problem is that we have to pick the thief's disguise for next year's vaccine nine months in advance. If we guess wrong, the vaccine doesn't work.

This paper introduces a new way to predict the thief's next move by looking at patterns of repetition, rather than just guessing. Here is the breakdown using simple analogies:

1. The "Copycat" Detective Work

The scientists realized that when the flu virus finds a disguise that works really well (one that lets it escape our immune system), it doesn't just happen once. It happens over and over again, independently, in different parts of the world.

• The Analogy: Imagine you are a detective trying to predict which new fashion trend will take over the city. You don't just look at one person wearing a hat. You look for convergence: if you see 50 different people in 50 different cities all suddenly buying the exact same weird hat, you know that hat is going to be huge.
• The Science: The researchers built a massive family tree (phylogeny) of the flu virus. They counted how many times specific genetic "typos" (mutations) appeared on different branches of the tree. If a specific typo keeps popping up in unrelated branches, it's a "copycat" signal. It means nature is selecting that specific change because it gives the virus a superpower.

2. The "Seven Magic Keys"

The virus has a huge surface with thousands of spots where it could potentially change. But the scientists found that the virus almost always changes in just seven specific spots near the "keyhole" (the receptor binding site) to break the lock.

• The Analogy: Think of the virus as a giant safe with a million dials. Most people think you have to guess the right number on any dial. But this paper shows that the safe only opens if you turn seven specific dials. If you see someone turning one of those seven dials, you know they are trying to crack the safe.
• The Science: By focusing only on these seven critical spots, the researchers could ignore the "noise" of the rest of the virus and focus on the changes that actually matter for vaccine matching.

3. The "Crystal Ball" (Predicting the Future)

The biggest breakthrough is timing. Usually, we only know a new flu strain is dominant after it has already taken over the world. This method allows us to see the signal before the virus becomes dominant.

• The Analogy: Imagine a new song is about to become a global hit. You don't wait until it's #1 on the charts to know it's a hit. You notice it first in a few small coffee shops, then in a few radio stations in different cities. If you see the song being played in 10 different cities by 10 different DJs, you can predict it will be a hit before it hits #1.
• The Science: The researchers found that the "copycat" signal (the virus trying the same mutation in different lineages) appears more than a year before that mutation takes over the global population. This gives vaccine makers a huge head start.

4. The "Traffic Light" System

The paper developed a scoring system called the Log Convergence Ratio (LCR).

• Green Light (High Score): The mutation is appearing everywhere independently. It's a winner. It's likely the next big thing.
• Red Light (Low Score): The mutation is rare or only appearing by accident. It's a dead end.
• Yellow Light: It's appearing, but maybe not enough to be sure yet.

Why This Matters

Currently, picking the flu vaccine is a bit like guessing which team will win the Super Bowl based on last year's stats. Sometimes we get it right, but often we get it wrong, and the vaccine is less effective.

This new method is like having a super-advanced scouting report. By watching for the "copycat" patterns in the seven key spots, scientists can tell the World Health Organization: "Hey, the virus is trying out this specific disguise in multiple places. It's going to be the winner next year. Let's build the vaccine against that one."

The Bottom Line

The virus is trying to evolve, but it's not as random as we thought. It keeps hitting the same few buttons to win. By counting how often those buttons are hit by different "copycats," we can predict the virus's next move years in advance, ensuring our vaccines are ready before the virus even arrives.

Technical Summary

1. Problem Statement

Seasonal influenza vaccination relies on selecting vaccine strains approximately 9 months before the flu season. However, influenza viruses, particularly the H3N2 subtype, undergo rapid antigenic evolution (antigenic drift) driven by mutations in the Hemagglutinin (HA) protein. This often leads to a mismatch between the vaccine strain and circulating viruses, reducing vaccine effectiveness.

Current predictive methods (based on variant frequencies, phylogenetic tree shape, or long-term dN/dS ratios) struggle to predict the emergence of specific future antigenic variants before they become dominant. There is a critical need for a method that can prospectively identify which specific amino acid substitutions are likely to drive the next major antigenic cluster transition, ideally well before the new cluster establishes itself in the population.

2. Methodology

The authors developed a novel metric called the Log Convergence Ratio (LCR) to estimate the fitness effects of individual amino acid substitutions by measuring convergent evolution in a global phylogeny.

• Data Source: A large phylogenetic tree constructed from all available human H3N2 HA1 nucleotide sequences from GISAID (up to January 2025).
• Tree Construction: Maximum-likelihood trees were built using IQTREE, CMAPLE, and UShER. Branches were annotated with nucleotide mutations.
• The LCR Metric:
  • Concept: Adaptive substitutions often arise independently multiple times (convergent evolution) in different lineages. The frequency of these independent occurrences relative to a neutral expectation indicates positive selection.
  • Calculation:
    1. Count the observed number of independent occurrences ($n_{obs}$) of a specific amino acid substitution across the tree.
    2. Calculate the neutral expected number of occurrences ($n_{exp}$). This is derived by modeling the mutation rates of synonymous nucleotide changes at four-fold degenerate sites. Crucially, the model accounts for the fact that different nucleotide mutations (e.g., GA vs. CG) occur at vastly different rates (up to 50x difference).
    3. Compute the Convergence Ratio: $CR = n_{obs} / n_{exp}$.
    4. Compute the Log Convergence Ratio: $LCR = \log_2(CR)$.
  • Significance Testing: A null model incorporating Poisson-distributed sampling variation and log-normally distributed variation in synonymous mutation rates was used to calculate empirical p-values.
• Target Positions: The analysis focused on seven key amino acid positions near the HA receptor binding site (positions 145, 155, 156, 158, 159, 189, and 193), previously identified as the primary drivers of major antigenic cluster transitions.

3. Key Contributions

• Substitution-Level Resolution: Unlike traditional dN/dS methods that estimate selection at the site (position) level, this method estimates selection for specific amino acid substitutions. This allows for the differentiation between beneficial and deleterious changes at the same position.
• Prospective Prediction: The study demonstrates that signals of positive selection (high LCR) for future antigenic transitions can be detected more than one year before the new antigenic cluster becomes dominant in the population.
• Integration of Mutational Bias: The method explicitly corrects for nucleotide mutation rate biases (e.g., transition vs. transversion rates), which is critical for accurate fitness estimation.
• Real-World Application: The authors applied this method to generate shortlists of high-risk substitutions for the World Health Organization (WHO) Vaccine Composition Meetings, successfully identifying emerging lineages (e.g., the 158K/189R double mutant) before they became dominant.

4. Key Results

• Validation of Selection Signals:
  • Substitutions on the "trunk" of the phylogenetic tree (leading to future dominant strains) and those leading to new antigenic clusters are typically under positive selection (positive LCR).
  • Substitutions on minor branches are typically neutral or deleterious.
• Predictive Power:
  • For 12 out of 14 major antigenic cluster transitions since 1987, the causative substitution was ranked in the top 6 (median rank: 3rd) among all possible substitutions at the seven key positions based on LCR calculated before the new cluster emerged.
  • The signal remained robust even when calculated using sequences collected >1 year before the new cluster reached 5% frequency.
• Case Studies:
  • F159Y (PE09 to HK14): Ranked 2nd in the ancestral PE09 cluster before the transition.
  • F159S (PE09 to SW13): The only case where the transition substitution had a negative LCR (indicating negative selection). This lineage (SW13) eventually went extinct, outcompeted by the HK14 lineage (driven by F159Y), confirming the LCR's accuracy in identifying the fitter path.
  • Recent Success (DA21): In the currently circulating DA21 cluster, the method identified S145N, N158K, and K189R as having high LCRs. These substitutions were tracked in real-time, leading to the identification of a "double mutant" clade (158K+189R) that rapidly spread globally in 2025, prompting updated WHO vaccine recommendations.
• Limitations & Nuances:
  • Epistasis: Some substitutions (e.g., Y159N) showed noisy signals due to epistatic interactions with other mutations (e.g., Y195F).
  • Sampling Bias: The method relies on sufficient sequencing depth. Substitutions caused by rare nucleotide mutations (e.g., CG) are harder to detect but have historically never caused major trunk transitions.
  • Multi-nucleotide changes: The method currently focuses on single-nucleotide substitutions; multi-nucleotide changes are harder to detect with current sequencing depths.

5. Significance

This paper presents a paradigm shift in influenza surveillance and vaccine strain selection. By quantifying the "convergent fitness" of specific mutations, the LCR metric provides an early warning system for antigenic drift.

• Improved Vaccine Matching: The ability to predict antigenic transitions >1 year in advance allows for more proactive vaccine strain selection, potentially reducing the frequency of vaccine mismatches.
• Generalizability: The methodology is not limited to influenza; it can be applied to other rapidly evolving pathogens (like SARS-CoV-2) to predict evolutionary trajectories based on phylogenetic convergence.
• Operational Impact: The study demonstrates that this approach is already being used to inform WHO recommendations, bridging the gap between theoretical evolutionary biology and public health policy.

In summary, the authors have successfully translated the concept of convergent evolution into a robust, quantitative tool for forecasting the antigenic future of seasonal influenza, offering a significant improvement over existing frequency-based or tree-shape-based prediction models.