Source-linked AI summary
Integrating influenza antigenic dynamics with molecular evolution
Trevor Bedford, Marc A. Suchard, Philippe Lemey, Gytis Dudas, Victoria Gregory, Alan J. Hay, John W. McCauley, Colin A. Russell, Derek J. Smith, Andrew Rambaut
TL;DR
Influenza antigenic evolution allows immune escape, but antigenic and genetic evolution have often been analyzed separately. The paper models antigenic diffusion over a shared virus phylogeny using HI data from four influenza lineages. It finds that A/H3N2 evolves faster and more punctuatedly than other lineages, while antigenic drift appears to drive strain replacement and seasonal incidence patterns.
Problem
Influenza antigenic drift enables repeated infection and reduces protection from fixed vaccines, while antigenic and genetic evolution have commonly been analyzed in isolation.
Method
The study uses a flexible Bayesian framework to jointly model antigenic phenotype and genetic evolution over a shared virus phylogeny, using HI data from four influenza lineages.
Results
A/H3N2 evolves faster and in a more punctuated fashion than other influenza lineages, and antigenic drift appears to drive strain replacement and contribute to seasonal incidence patterns.
Takeaways & Limitations
The framework provides a foundation for evolutionary antigenic cartography that simultaneously assesses antigenic phenotype and its evolution.
Takeaways & Limitations
HI assays have limited sensitivity for temporally distant viruses, leaving threshold titers that make global antigenic trajectories difficult to distinguish.
Abstract
from arXiv · showhide
Influenza viruses undergo continual antigenic evolution allowing mutant viruses to evade host immunity acquired to previous virus strains. Antigenic phenotype is often assessed through pairwise measurement of cross-reactivity between influenza strains using the hemagglutination inhibition (HI) assay. Here, we extend previous approaches to antigenic cartography, and simultaneously characterize antigenic and genetic evolution by modeling the diffusion of antigenic phenotype over a shared virus phylogeny. Using HI data from influenza lineages A/H3N2, A/H1N1, B/Victoria and B/Yamagata, we determine patterns of antigenic drift across viral lineages, showing that A/H3N2 evolves faster and in a more punctuated fashion than other influenza lineages. We also show that year-to-year antigenic drift appears to drive incidence patterns within each influenza lineage. This work makes possible substantial future advances in investigating the dynamics of influenza and other antigenically-variable pathogens by providing a model that intimately combines molecular and antigenic evolution.
Introduction
Seasonal influenza continually evolves antigenically, enabling repeated infections and weakening protection from fixed vaccines. The paper addresses this by combining antigenic measurements with genetic relationships to study evolution across major influenza lineages.
- Seasonal influenza infects 10% to 20% of people annually and causes an estimated 250,000 to 500,000 deaths.
- Antigenic mutations reduce immune recognition, allowing individuals to be infected repeatedly and causing fixed vaccine efficacy to decline over time.
- The four major human influenza lineages are A/H3N2, A/H1N1, B/Victoria, and B/Yamagata.
- The HI assay measures cross-reactivity between virus strains and sera, while antigenic maps quantify similarity and distance from many HI measurements.
- Previous analyses treated antigenic maps and genetic phylogenies largely separately, although they often contain many of the same isolates.
- The study jointly characterizes antigenic and genetic evolution and reconstructs antigenic dynamics for all four circulating influenza lineages.
Results and discussion
The Bayesian framework combines antigenic cartography with phylogenetic evolution, revealing lineage-specific drift patterns and their relationship to influenza incidence. A/H3N2 evolves fastest and most punctuately, while genetic distance offers only limited prediction of antigenic distance.
- Model performance: A two-dimensional BMDS model performed well across lineages, with average absolute predictive error of 0.78–0.91 log2 HI titers.The model was therefore used in subsequent analyses.
- Genetic–antigenic relationships: Genetic and phylogenetic distances correlate with antigenic distance, but low R2 values show that genetics alone incompletely resolves antigenic differences.The shared phylogenetic diffusion prior incorporates genetic similarity while allowing antigenic drift and phylogenetic covariance.
- Lineage comparisons: A/H3N2 shows substantially more antigenic evolution than A/H1N1, B/Vic, or B/Yam, with prominent clusters and some evolutionary dead-ends.The analysis reconstructs antigenic and evolutionary relationships across all four lineages.
- Limitations: HI assay sensitivity limits long-range A/H3N2 comparisons and makes linear versus slightly curved antigenic trajectories difficult to distinguish.The analysis therefore interprets map locations locally and avoids strong claims about the larger configuration.
- Lineage comparisons: A/H3N2 drifts at 1.01 antigenic units per year and exhibits occasional large jumps associated with cluster transitions.Most variation occurs along antigenic dimension 1; dimension 2 varies when distinct lineages coexist transiently.
- Lineage comparisons: A/H1N1, B/Vic, and B/Yam drift at 0.62, 0.42, and 0.32 units per year, respectively, with punctuated evolution less evident in the B lineages.All four lineages nevertheless show similar levels of standing antigenic variation because diffusion volatility varies less than drift rates.
- Epidemiological relationship: Years with pronounced antigenic drift tend to show increased incidence within each lineage, with Pearson correlations of 0.51, 0.29, 0.44, and 0.14.Bootstrap p-values were 0.056, 0.201, 0.097, and 0.341, respectively; isolate-count analysis found little correlation with drift estimates (r = −0.01).
Conclusions
The study establishes evolutionary antigenic cartography as a framework for jointly analyzing antigenic phenotype and viral evolution across four influenza lineages. It finds lineage-specific differences in antigenic drift and links antigenic evolution with incidence and future modeling opportunities.
- The framework simultaneously characterizes antigenic phenotype and evolutionary relationships across A/H3N2, A/H1N1, B/Vic and B/Yam.
- A/H3N2 evolves antigenically faster than A/H1N1, which evolves faster than B/Vic or B/Yam.
- Antigenic evolution within each lineage drives strain replacement and contributes to seasonal incidence patterns.
- The framework can incorporate assay covariates and estimate internal-node sequence states and antigenic locations to connect substitutions with antigenic diffusion.
- Combining evolutionary and antigenic information may help identify low-frequency, expanding antigenically novel lineages for vaccine strain selection.
Materials and methods
The paper extends antigenic cartography with Bayesian multidimensional scaling that probabilistically models HI measurements and antigenic locations. It represents viruses and sera in a low-dimensional map while accommodating threshold and interval observations.
- HI assays measure cross-reactivity between virus strains and antibodies as titers representing serum dilution before inhibition ceases.
- Titers are analyzed in log2 space because experimental measurements typically follow twofold dilution series, producing a sparse observation matrix H.
- Traditional MDS optimized cartographic locations but did not integrate genetic data, motivating the Bayesian extension.
- BMDS places viruses and sera in a low-dimensional antigenic map, with distances inversely proportional to cross-reactivity and typically P = 2.One antigenic unit corresponds to an expected 2-fold drop in HI titer.
- The model represents virus and serum locations as coordinate matrices X and Y, with distances defined by their L2 norm.
- Likelihoods accommodate exact, threshold, and interval HI measurements, and analyses use interval likelihoods rather than point likelihoods unless otherwise noted.
- The overall likelihood multiplies probabilities for individual measurements, using diffuse normal priors on virus and serum locations.
Virus avidity and serum potency
The model addresses assay variation by allowing serum potency and virus avidity to vary rather than treating all maximum titers and isolates as equally reactive.
- Serum potency: Serum potency is modeled as a random quantity because measurements against only distant viruses can make a serum’s maximum titer appear artificially low.This relaxes the assumption of fixed serum-specific values.
- Virus avidity: Virus avidity represents each isolate’s general reactivity level across hemagglutination inhibition assays.Including avidity accounts for systematic differences among virus isolates.
- Virus avidity: Virus avidities are estimated hierarchically using the empirical mean and variance of viruses’ maximum observed titers.The reference values are derived from each virus’s maximum HI titer across sera.
Drift model of antigenic evolution
The drift model represents virus and serum locations as moving through antigenic space over time, while accounting for uncertainty in their temporal distributions.
- Model identifiability: Multiple virus and serum configurations can produce the same observed HI-data likelihood, preventing absolute antigenic locations from being identified.The data may distinguish relative but not absolute positions on the antigenic map.
- Temporal drift: The time variable t measures the interval from an indexed virus or serum to the earliest sampled virus or serum.This temporal reference anchors the drift model to the earliest sampled observation.
- Temporal drift: Virus and serum locations drift in a line across the antigenic map at rate µ, with σx and σy controlling the breadth of virus and serum location clouds.σx describes virus-location spread, whereas σy describes serum-location spread.
Phylogenetic diffusion model of antigenic evolution
The phylogenetic diffusion model couples antigenic locations to viral evolutionary history by letting shared ancestry induce covariance among virus positions.
- Phylogenetic coupling: Virus locations are modeled with a Brownian-motion prior whose covariance reflects shared evolutionary history.This replaces a prior of independent virus locations and simultaneously models antigenic locations and genetic relatedness.
- Phylogenetic diffusion: The phylogeny τ induces genetically similar viruses to have nearby prior antigenic locations, with σx controlling Brownian-motion volatility.Phylogeny tips correspond to observed virus locations, while internal states determine their joint probability.
- Phylogenetic diffusion: The process is a Wiener process with drift, where µ changes expected states along the phylogeny while diffusion determines location variability.The root location follows a normal distribution, and internal phylogenetic states are integrated analytically.
- Phylogenetic inference: The virus phylogeny τ is estimated from sequence data using established methods implemented in BEAST.The resulting tree supplies the evolutionary structure used by the antigenic diffusion model.
Posterior inference
The model uses diffuse Gamma priors and samples a posterior combining immunological data with virus and serum locations and other parameters. MCMC procedures in BEAST explore location, avidity, potency, scale, drift, precision, and phylogenetic-tree parameters.
- Diffuse Gamma(a, b) priors with a = 0.001 and b = 0.001 provide little-to-no weight on the posterior distributions.
- The posterior factors the likelihood of immunological data and priors for virus locations, serum locations, and model parameters.
- MCMC proposals translate or scale virus and serum locations and update virus avidities, serum potencies, and model-scale parameters.
Genetic, antigenic and surveillance data
The study assembled HI measurements, sequence data, and surveillance records across four human influenza lineages. These datasets combine published and vaccine-selection sources with CDC seasonal reports.
- A/H3N2 antigenic data were compiled from prior publications, vaccine strain-selection reports, and database-matched HA nucleotide sequences.
- 115 A/H1N1 virus isolates, 77 serum isolates, and 1882 HI measurements covered 1977–2009.
- 179 B/Victoria virus isolates, 70 serum isolates, and 2003 HI measurements covered 1986–2011.
- 174 B/Yamagata virus isolates, 69 serum isolates, and 1962 HI measurements covered 1987–2011.
- Surveillance records came from CDC FluView yearly influenza-season summaries spanning 1997–1998 to 2010–2011.
Implementation
Phylogenetic trees were estimated with BEAST and then used alongside MCMC samples to infer antigenic locations and diffusion paths. Posterior summaries addressed parameter uncertainty despite alignment challenges among cartographic samples.
- Phylogenetic trees for all four lineages used BEAST, the SRD06 substitution model, a constant-size coalescent, and a strict molecular clock.MCMC produced 2000 sampled trees after burn-in.
- The antigenic model sampled virus and serum locations, avidities, potencies, precision parameters, drift rate, and the phylogenetic tree using MCMC.
- Including drift parameter µ corrected most identifiability issues and substantially improved mixing of antigenic locations.
- Local isometries can rotate antigenic clusters across posterior samples, making full alignment with Procrustes analysis difficult.
- Diffusion paths were reconstructed backward through the phylogeny using sampled tip locations and peeling-based internal-node locations.The procedure yielded 2000 posterior trees tagged with estimated tip and internal-node locations.
Comparison with previous results
The BMDS comparisons reproduce local antigenic relationships found by MDS while revealing global flexibility and addressing identifiability through drift-informed priors. Estimating avidities and potencies also changes map geometry while preserving local consistency.
- Comparison with previous results: 273 virus isolates, 79 serum isolates, and 4252 HI measurements from 1968–2003 were used to compare BMDS with Smith et al.’s MDS results.
- Comparison with previous results: Multiple BMDS solutions had highly similar likelihoods, including alternatives involving rotation of the HK/68, EN/72, and VI/75 clusters.
- Comparison with previous results: Viruses up to approximately 15 years divergent showed locally consistent solutions despite global flexibility in antigenic locations.
- Comparison with previous results: Including a drift prior removed identifiability problems, preserved local MDS consistency, and partitioned more variance to the first antigenic dimension.
- Comparison with previous results: The drift prior slightly improved test error for A/H1N1, B/Victoria, and B/Yamagata datasets.
- Comparison with previous results: The final BMDS model incorporated temporally and phylogenetically informed priors on antigenic locations and estimated serum and virus avidities.
- Comparison with previous results: Estimating avidity and potency produced a more linear map, while preserving local consistency for viruses less than approximately 10 years divergent.A/Bilthoven/16190/1968 became more distant from other viruses under this specification.
- Comparison with previous results: For viruses at least 15 years divergent, threshold titers require indirect inference of relative locations, limiting global comparisons.Local comparisons used to calculate year-to-year drift are expected to remain robust to many model particulars.
Availability
The cartographic model code and the incidence and HI datasets used in the analysis are publicly archived or downloadable.
- The cartographic models’ source code is available through the BEAST software package and its Google Code repository.Additional implementation examples are available in the flux GitHub repository.
- The incidence and HI data used in the analysis are archived with Dryad.The archive is identified by DOI 10.5061/dryad.rc515.
Funding
The research received support from individual fellowships, government grants, foundation grants, and European Research Council funding.
- TB was supported by a Newton International Fellowship from the Royal Society.
- MAS was supported by NIH grants R01 HG006139 and R01 AI107034 and NSF grants DMS126153 and IIS1251151.
- The research received European Research Council funding through FP7 agreements 278433-PREDEMICS and 260864.