Real-time tracking and prediction of influenza virus evolution

Trevor Bedford (@trvrb)
25 May 2017
Immunology and Evolution of Influenza Symposium
Emory University

Want to forecast the make up of the future flu population from the population that exists today

Real-time updates as new information rolls in

Population turnover (in H3N2) is extremely rapid

Clades emerge, die out and take over

Clades show rapid turnover

Dynamics driven by antigenic drift

Drift variants emerge and rapidly take over in the virus population

This causes the side effect of evading existing vaccine formulations

Drift necessitates vaccine updates

H3N2 vaccine updates occur every ~2 years

Timely surveillance and rapid analysis essential to vaccine strain selection


Project to provide a real-time view of the evolving influenza population


Project to provide a real-time view of the evolving influenza population

All in collaboration with Richard Neher

nextflu pipeline

  1. Download all recent HA sequences from GISAID
  2. Filter to remove outliers
  3. Subsample across time and space
  4. Align sequences
  5. Build tree
  6. Estimate clade frequencies
  7. Infer antigenic phenotypes
  8. Export for visualization

Up-to-date analysis publicly available at:

Antigenic analysis

Influenza hemagglutination inhibition (HI) assay

HI measures cross-reactivity across viruses

Data in the form of table of maximum inhibitory titers

Antigenic cartography compresses HI measurements into an interpretable diagram

Instead of a geometric model, we sought a phylogenetic model of HI titer data

Identify phylogeny branches associated with drops in HI titer

Model can be used to interpolate across tree and predict phenotype of untested viruses

Model is highly predictive of missing titer values

Recent HI data from London WHO Collaborating Center

Up-to-date analysis at:

Mapping genotype to phenotype changes

Proliferation of "epitope masks" in the literature

Inferences of branch-specific amino acid changes and branch-specific phenotype changes

Can start to validate these masks

Huddleston 2016

And infer masks that better map to HI data

Huddleston 2016


"The future is here, it's just not evenly
distributed yet"
— William Gibson

USA music industry, 2011 dollars per capita

Influenza population turnover

Vaccine strain selection timeline

Seek to explain change in clade frequencies over 1 year

Fitness models can project clade frequencies

Clade frequencies $X$ derive from the fitnesses $f$ and frequencies $x$ of constituent viruses, such that

$$\hat{X}_v(t+\Delta t) = \sum_{i:v} x_i(t) \, \mathrm{exp}(f_i \, \Delta t)$$

This captures clonal interference between competing lineages

The question of forecasting becomes: how do we accurately estimate fitnesses of circulating viruses?

Fortunately, there's lots of training data and previously successful strains have had:

  1. Amino acid changes at epitope sites
  2. Antigenic novelty based on HI
  3. Rapid phylogenetic growth

Predictor: calculate HI drop from ancestor,
drifted clades have high fitness

Predictor: project frequencies forward,
growing clades have high fitness

We predict fitness based on a simple formula

where the fitness $f$ of virus $i$ is estimated as

$$\hat{f}_i = \beta^\mathrm{HI} \, f_i^\mathrm{HI} + \beta^\mathrm{freq} \, f_i^\mathrm{freq}$$

where $f_i^\mathrm{HI}$ measures antigenic drift via HI and $f_i^\mathrm{freq}$ measures clade growth/decline

We learn coefficients and validate model based on previous 15 H3N2 seasons

Clade growth rate is well predicted (ρ = 0.66)

Growth vs decline correct in 84% of cases

Trajectories show more detailed congruence

Trajectories show more detailed congruence

This model is similar in formulation and performance to Łuksza and Lässig

We find that HI is preferred over epitope mutations

Model Ep coefficient HI coefficient Freq error Growth corr
Epitope only 2.36 -- 0.10 0.57
HI only -- 2.05 0.08 0.63
Epitope + HI -0.11 2.15 0.08 0.67

When does the forecast fail?

Emerging clades are difficult to forecast: little antigenic data and little evidence of "past performance"

Models work well for clades at >10%, but less well for clades <5%

New mutations difficult

Models can project forward from circulating strains, but cannot foresee the appearance of new mutations

Intrinsically limits the timescale of forecasting to ~1 year

Model is only as good as the data

Requires rapid shipping of samples, rapid sequencing and rapid antigenic characterization

Current situation

Looking forward

Further improvements to predictive modeling

  1. Extend to other seasonal viruses
  2. Forecast NA evolution
  3. Integrate neutralization (FRA) assay data
  4. Model effects of egg adaptation
  5. Incorporate an explicit geographic model

Phylogeny of H3 with geographic history

Geographic location of phylogeny trunk


Analysis: Richard Neher, Colin Russell, Charlton Callender, John Huddleston, Gytis Dudas, Colin Megill, Andrew Rambaut, Charles Cheung, Marc Suchard, Steven Riley, Philippe Lemey, Boris Shraiman, Marta Łuksza, Michael Lässig

GISRS: Ian Barr, Shobha Broor, Mandeep Chadha, Nancy Cox, Rod Daniels, Becky Garten, Palani Gunasekaran, Aeron Hurt, Anne Kelso, Jackie Katz, Nicola Lewis, Xiyan Li, John McCauley, Takato Odagiri, Varsha Potdar, Yuelong Shu, Eugene Skepner, Masato Tashiro, Dayan Wang, Dave Wentworth, Xiyan Xu