Interpreting viral evolution


Trevor Bedford

Fred Hutchinson Cancer Center / Howard Hughes Medical Institute
11 Jul 2025
SURP Seminar
Fred Hutch
Slides at: bedford.io/talks

We work at the interface of virology, evolution and epidemiology

Ebola
Zika

Segue through data vizualization

Data tells a story

Charles Minard and Napoleon's 1812 campaign

If data are looked at in the right way, answers become immediately clear

Edward Tufte

Above all, show the data

A good graphic is honest

Barchart of average value

Box-and-whisker plot showing variation

Showing the data

To clarify add detail

"To clarify, add detail... Clutter and overload are not attributes of information, they are failures of design. If the information is in chaos, don’t start throwing out information, instead fix the design."
– Tufte

Evolution of Conus shells

Greenland rising

Small multiples and dimensionality

Mapping infection spread

John Snow and the founding of epidemiology

My research

Methods focus on sequencing to reconstruct pathogen spread

Epidemic process

Sample some individuals

Sequence and determine phylogeny

Sequence and determine phylogeny

Pathogen genomes may reveal

  • Evolution of new adaptive variants
  • Epidemic origins
  • Patterns of geographic spread
  • Animal-to-human spillover
  • Transmission chains

Influenza: Forecasting spread of new variants for vaccine strain selection

Zika: Uncovering origins of the epidemic in the Americas

Ebola: Revealing spatial spread and persistence in West Africa

MERS: Repeated spillover into the human population from camel reservoir

Actionable inferences

Nextstrain

Project to conduct real-time genomic epidemiology and evolutionary analysis of emerging epidemics

Richard Neher,   Ivan Aksamentov,   John SJ Anderson,   Kim Andrews,   Jennifer Chang   James Hadfield,   Emma Hodcroft,   John Huddleston,   Jover Lee,   Victor Lin,   Cornelius Roemer,   Thomas Sibley

Nextstrain architecture

Central aims: (1) rapid and flexible bioinformatic workflows, (2) interactive visualization and (3) always up-to-date analyses at nextstrain.org

All code open source at github.com/nextstrain

Operationalized during the 2018-2020 outbreak of Ebola in the Democratic Republic of the Congo

Genomic epidemiology during the COVID-19 pandemic

To date, over 17M SARS-CoV-2 genomes shared to GISAID and evolution tracked in real-time at nextstrain.org

Three key insights that genomic epi provided during pandemic

  1. Rapid human-to-human spread in Wuhan beyond initial market outbreak
  2. Extensive local transmission while testing was rare
  3. Identification of variants of concern and mapping of increased transmission rates

On Jan 19, 2020, the first 13 viral genomes from Wuhan and Bangkok showed little genetic diversity

Allowed us to infer introduction between Nov 15 and Dec 15 and subsequent rapid human-to-human spread

Rapid global epidemic spread from China

Epidemic in the USA was introduced from China in late Jan and from Europe during Feb

Early sequencing provided best estimate of extent of local outbreak

After initial wave, with mitigation
efforts and decreased travel,
regional clades emerge

Emergence of Alpha in the UK with excess spike mutations

Further emergence of variants with increased transmissibility

Variant emergence and spread has continued

Rapid displacement of existing diversity by emerging variants that escape from existing population immunity

Mutations in S1 domain of spike protein driving displacement

SARS-CoV-2 evolution fast relative to previous endemic viruses

Current directions

Wealth of genome data represents a foundational resource for tracking viral evolution and epidemiology

Though we're still only scratching the surface of this data, in large part due to methodological limitations

AI approaches

  • New approaches follow recent human language AI models and treat DNA nucleotides or protein amino acids as a text corpus
  • These models follow the familiar generative logic of being trained to predict the next token (nucleotide / amino acid)
  • The resulting models have a rich biological understanding encoded into the resulting internal model representations

Large language models are based on a transformer architecture

  • Models are trained to predict the next token from the preceeding tokens
  • Words (or parts of words) are converted into input embedding vectors and these vectors are multiplied through a large stack of matrices
  • The resulting output embedding vectors determine predicted token probabilities

Sequence language models have amazing potential

Evo 2 is trained on 9.3 trillion DNA bases and predicts mutation effects, learns biological features and can generate quasi-life-like sequences

Evo 2 predicts fitness effects of genomic changes

Under the hood, biological features inform Evo 2 embeddings

Acknowledgements

Influenza: WHO Global Influenza Surveillance Network, GISAID, Richard Neher, John Huddleston.   Zika: Nick Loman, Nuno Faria, Oli Pybus, Josh Quick, Ingra Claro, Julien Thézé, Jaquilene de Jesus, Marta Giovanetti, Moritz Kraemer, Sarah Hill, Ester Sabino, Luiz Alcantara, Allison Black.   Ebola: Gytis Dudas, Andrew Rambaut, Luiz Carvalho, Philippe Lemey, Marc Suchard, Andrew Tatem, Eddy Kinganda-Lusamaki, Placide Mbala-Kingebeni, Catherine Pratt, Jean-Jacques Muyembe Tamfum.   MERS: Gytis Dudas, Andrew Rambaut, Luiz Carvalho.   SARS-CoV-2: Data producers from all over the world, GISAID, the Nextstrain team, Marlin Figgins, Katie Kistler.   Nextstrain: Richard Neher, Ivan Aksamentov, John SJ Anderson, Kim Andrews, Jennifer Chang, James Hadfield, Emma Hodcroft, John Huddleston, Jover Lee, Victor Lin, Colin Megill, Cornelius Roemer, Thomas Sibley.

Bedford Lab: John Huddleston,   James Hadfield,   Katie Kistler,   Thomas Sibley,   Jover Lee,   Marlin Figgins,   Victor Lin,   Jennifer Chang,   Nashwa Ahmed,   Cécile Tran Kiem,   Kim Andrews,   Philippa Steinberg,   Jacob Dodds,   John SJ Anderson,   Nobuaki Masaki,   Amin Bemanian,   Carlos Avendano,   Aayush Verma,   Lucy O'Brien