EP.002 · Research talk · Science 2022

Which Mutations Make a Virus Spread? Reading 6.4 Million Genomes

Dr. Jacob E. Lemieux · Infectious disease physician-scientist, Massachusetts General Hospital

Hosted by Dr. Omar Costilla Reyes

First published On Interpretable Updated 59 minAI in HealthPublic HealthInfectious DiseaseGenomicsInterpretable AI

Listen or watch

Download MP3(21 MB)
0:0059:14

Audio edition, 59 min. Keyboard: K play/pause, J back 15 s, L forward 30 s.

Full video, 59 min.

Chapters(17)

About this episode

New SARS-CoV-2 variants kept emerging, and public health teams needed to know quickly which ones were likely to take over. In this talk, infectious disease physician-scientist Jacob Lemieux walks through PyR0, a hierarchical Bayesian model written in the Pyro probabilistic programming language and fit to the publicly available SARS-CoV-2 genomes of the time. The model estimates the relative fitness of viral lineages across regions, detects lineages that are rising in prevalence, and identifies mutations, inside and outside the spike protein, that are associated with increased fitness. The work, led by Fritz Obermeyer with colleagues at the Broad Institute of MIT and Harvard and Massachusetts General Hospital, was published in Science in 2022. That structure is why it belongs on Interpretable: instead of a black-box forecast, PyR0 reports how much each mutation contributes to a lineage's growth, with uncertainty, in terms virologists and public health teams can check against laboratory experiments.

Listen if: You work in public health, epidemiology, or genomics, or want to see what an interpretable model looks like at the scale of millions of sequences.

The research behind the episode

Science2022Peer reviewed

Analysis of 6.4 million SARS-CoV-2 genomes identifies mutations associated with fitness

Fritz Obermeyer, Martin Jankowiak, Nikolaos Barkas, Stephen F. Schaffner, et al.

376 (6599): 1327–1332 · doi:10.1126/science.abm1208

In plain language

PyR0 is a hierarchical Bayesian model that estimates how quickly each SARS-CoV-2 lineage grows relative to the others in every region, and attributes that growth to individual mutations. Fit to 6.4 million genomes, it ranks lineages as new sequences arrive and forecasts which new lineages are likely to spread.

Why it matters for health

Genomic surveillance helps only when it is fast and readable. A fitness estimate for each mutation tells public health teams and laboratories where to look next.

Key findings

  1. 01Many mutations increase fitness, including known spike mutations and many outside the spike, in the nucleocapsid and nonstructural proteins.
  2. 02The growth of a new lineage can be forecast from its mutational profile.
  3. 03The model prioritizes mutations of biological and public health concern for laboratory follow-up.

What you'll learn

  • Why fast-spreading variants follow logistic growth curves, and how to model every lineage at once
  • How PyR0 links individual mutations to lineage fitness and forecasts growth from a mutational profile
  • Where genomic surveillance models fall short, and how they help prioritize mutations for lab follow-up

Working on something like this?

Suggest a guest or a paper for a future episode, or talk to equ about AI for health in your own work.