Frequency dynamics predict viral fitness, antigenic relationships and epidemic growth
Curation statements for this article:-
Curated by eLife
eLife Assessment
In this important work, the authors develop methods to forecast epidemic growth from viral sequencing data alone. The evidence for the usefulness of the approach is solid, but some justifications and methodological details are incomplete. This study should be of broad interest to the community interested in viral dynamics and epidemiology.
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (eLife)
Abstract
During the COVID-19 pandemic, SARS-CoV-2 variants drove large waves of infections, fueled by increased transmissibility and immune escape. Current models focus on changes in variant frequencies without linking them to underlying transmission mechanisms of intrinsic transmissibility and immune escape. We introduce a frame-work connecting variant dynamics to these mechanisms, showing how host population immunity interacts with viral transmissibility and immune escape to determine relative variant fitness. We advance a selective pressure metric that provides an early signal of epidemic growth using genetic data alone, crucial with current underreporting of cases. Additionally, we show that a latent immunity space model approximates immunological distances, offering insights into population susceptibility and immune evasion. These insights refine real-time forecasting and lay the groundwork for research into the interplay between viral genetics, immunity, and epidemic growth.
Article activity feed
-
eLife Assessment
In this important work, the authors develop methods to forecast epidemic growth from viral sequencing data alone. The evidence for the usefulness of the approach is solid, but some justifications and methodological details are incomplete. This study should be of broad interest to the community interested in viral dynamics and epidemiology.
-
Reviewer #1 (Public review):
Summary:
This paper develops a formalism for quantifying epidemic dynamics in terms of relative fitnesses of circulating variants, uses the formalism to elucidate fundamental tradeoffs of epidemics driven by variants with increased transmissibility versus immune escape capability, shows the formalism implies a natural quantity measuring the impact of selection on epidemic growth, and demonstrates that the formalism enables a decomposition of epidemic dynamics into circulation among different immunity groups. The relative fitness formalism enables these analyses to be performed with genetic sequence data only, a major benefit of the model given the relatively high availability of sequence data compared to other data streams such as case counts and titers.
Strengths:
Linking epidemic dynamics to pathogen …
Reviewer #1 (Public review):
Summary:
This paper develops a formalism for quantifying epidemic dynamics in terms of relative fitnesses of circulating variants, uses the formalism to elucidate fundamental tradeoffs of epidemics driven by variants with increased transmissibility versus immune escape capability, shows the formalism implies a natural quantity measuring the impact of selection on epidemic growth, and demonstrates that the formalism enables a decomposition of epidemic dynamics into circulation among different immunity groups. The relative fitness formalism enables these analyses to be performed with genetic sequence data only, a major benefit of the model given the relatively high availability of sequence data compared to other data streams such as case counts and titers.
Strengths:
Linking epidemic dynamics to pathogen evolution is a fundamental problem in studies of antigenically variable pathogens, with models of epidemic dynamics and immune-driven evolution going back decades in applications to respiratory pathogens such as influenza. The COVID-19 pandemic heightened the urgency for developing methods for quantifying epidemic growth in contexts where novel variants emerge, leading to differential susceptibility among individuals with diverse exposure histories with implications for vaccination strategies. Real-world data streams such as case counts and immunological measurements have a variety of shortcomings that pose major challenges for quantitative models aiming to inform policy. In recent years, genetic sequencing data has become widely available for pathogens including SARS-CoV-2 and influenza, allowing tracking of pathogen evolution at unprecedented detail in real time, yet biases in the collection of sequence data across different populations make connections between absolute epidemic size and variant frequencies from sequence data not immediately transparent.
This paper's contributions are exciting because they demonstrate new ways to link pathogen evolution and epidemic dynamics using very accessible data. From a theoretical perspective, the model is appealing because of its simple derivation in terms of compartmental models of epidemics, which are standard in the literature, and its clear extension to populations with heterogeneous immune histories. The latter extension leads directly to new methods for inferring immune groups with differential susceptibility to antigenically distinct variants in populations with heterogeneous immune histories without access to immunological data such as titers, an important advance given the wide applicability of quantification of antigenic relationships among variants in real populations.
Weaknesses:
While the demonstrated methods for forecasting short-term epidemic growth and for quantifying population immunity using sequence data are exciting as proofs of principle, the validation and statistical support provided in the analyses have drawbacks that are not fully addressed in the manuscript, weakening the evidence for the usefulness of the methods in their current form.
The analyses forecasting epidemic growth using Gaussian process models are justified using Pearson correlation coefficients whose values are extremely low for the test data period. The explanation given for this is that the case data used to validate the predictions has worse ascertainment over time, but it is not shown directly that the model may be working well despite the low correlations. Whereas, by eye, the predicted epidemic growth curves appear to capture features of the observed epidemic growth curves, the computed metrics don't support the claim of success of the predictions. Additionally, nearly all the model fits lack estimates of uncertainty, so it is not possible to discern the significance of departures between the model and data, or subtle differences in relative fitness calculations across geographies.
The analysis of latent pseudo-immune components also suffers drawbacks that render it more of an interesting proof of principle than a convincing tool for prediction at this point. In particular, in figures S18 and S19, metrics meant to quantify the statistical significance of the results show no difference from null models computed by permuting variants and their escape vectors, yet no interpretation is given for the lack of significance. Moreover, the model fits relating titer distance to pseudo escape distance seem unsuccessful for JN.1 infection and XBB infection histories, which is not adequately accounted for in the text, which cites just "weaker correlations" in these cohorts.
In several instances, the evidence for the new data analyses is weakened by a lack of clarity in the presentation of the technical details of the methods. For example, in the discussion of the Gaussian process models, it was not clear what features of the problem inform the choice of kernel (Matern 5/2), which hyperparameters were used, and how novel this use of Gaussian processes is. In the section describing methods for predicting epidemic growth rate from selective pressure, the discussion of the gradient boosting regressor model provided no intuition as to why this method performed better than the others tested or whether this was particularly important to the conclusions, and the lack of discussion of uncertainty or variability in the model predictions makes it difficult to assess the significance of the time series estimates alone. In the discussion of the latent immune factor model, the mismatch between the notation used in Equation 5 compared to that in Equation 18 made the derivations more difficult to follow. Subsequently, the explanation of the fitting of the pseudo-immune model left out details, such as an explicit definition of distance in pseudo-escape space, to what extent the group-level mean aggregated titer measurement captured features of the titer data (despite ignoring interindividual variability), and a thorough discussion of the successes and shortcomings of the fits in different scenarios. More explicit presentation of the mathematical choices going into the methods, sources and quantification of uncertainty, and cases where the model performs well or poorly could significantly bolster the case for the usefulness of sequence data in quantitatively predicting epidemic growth and antigenic relationships among variants in practice, in more general settings than those carried out here.
-
Reviewer #2 (Public review):
Summary:
The authors first introduce a framework to understand how different phenotypic drivers of viral evolution, i.e., changes in transmissibility versus immune escape, complicate epidemic forecasting using only genetic data. To overcome these complications, they advance an evolutionary "selective pressure" metric to predict population-wide epidemic growth from genetic data alone. Separately, they introduce a latent space model to infer a "pseudo" population immune structure from geographic variation in viral lineage dynamics, and find that the inferred pseudo-structure predicts human serological data.
Strengths:
This paper begins with a useful pedagogical exposition on the connection between fitness-driven frequency dynamics and underlying mechanisms of viral-immune co-evolution. A major contribution of …
Reviewer #2 (Public review):
Summary:
The authors first introduce a framework to understand how different phenotypic drivers of viral evolution, i.e., changes in transmissibility versus immune escape, complicate epidemic forecasting using only genetic data. To overcome these complications, they advance an evolutionary "selective pressure" metric to predict population-wide epidemic growth from genetic data alone. Separately, they introduce a latent space model to infer a "pseudo" population immune structure from geographic variation in viral lineage dynamics, and find that the inferred pseudo-structure predicts human serological data.
Strengths:
This paper begins with a useful pedagogical exposition on the connection between fitness-driven frequency dynamics and underlying mechanisms of viral-immune co-evolution. A major contribution of this paper - a method to infer variant-specific escape properties from geographically non-uniform variant frequency dynamics alone - is an interesting and potentially timely one, given the advance of sequencing-based surveillance.
Weaknesses:
The logical flow of the pedagogy part of the text works against the reader, which is problematic since it motivates the rest of the text. Moreover, some important modelling choices and procedures, particularly with respect to the selective pressure metric, are only cursorily described in the methods section. The lack of explanation and detail, especially relative to more simple choices that are seemingly motivated by the authors' own theory, makes it difficult to understand and therefore assess their validity and/or necessity.
-
Reviewer #3 (Public review):
Summary:
This study introduces a new analytical framework to analyze how viral variant frequencies change over time and in different locations. Two examples are given that demonstrate where this approach can be useful and where other approaches can be ambiguous in characterizing novel variants. The authors then demonstrate that the spatiotemporal dynamics of variant frequencies can be used to predict future epidemic growth rates and to investigate how variants differ in immune escape.
Strengths:
(1) Examples are provided that make the study accessible for a general audience.
(2) The authors demonstrate that their approach is predictive both of overall epidemic growth rates and immunological distance between variants.
(3) The approach introduced in this study can be readily applied to current and future …
Reviewer #3 (Public review):
Summary:
This study introduces a new analytical framework to analyze how viral variant frequencies change over time and in different locations. Two examples are given that demonstrate where this approach can be useful and where other approaches can be ambiguous in characterizing novel variants. The authors then demonstrate that the spatiotemporal dynamics of variant frequencies can be used to predict future epidemic growth rates and to investigate how variants differ in immune escape.
Strengths:
(1) Examples are provided that make the study accessible for a general audience.
(2) The authors demonstrate that their approach is predictive both of overall epidemic growth rates and immunological distance between variants.
(3) The approach introduced in this study can be readily applied to current and future epidemiological challenges that are similar to SARS-CoV-2 with respect to the relative evolutionary timescales wherever there is spatiotemporal heterogeneity in the susceptible population.
Weaknesses:
(1) The authors conclude their abstract claiming that their method provides an early signal of epidemic growth. Can this be quantified? Could the authors perform retrospective analyses for sequences available through various cutoff times, identify how early significant new variants are detected, and compare this to other detection methods?
(2) Analysis depicted in Figure 4 and Figure S9 could be explored further than speculatively attributing weak correlation to declining reporting rates for US states. Exploring how correlation between data and prediction varies over time during the test period might identify periods/events that explain weak correlation overall. The authors could explore predicting growth rates for estimated state prevalences rather than reported cases.
-
