Erik Volz

The fitness of a pathogen can be characterized at multiple levels corresponding to different life cycle stages. Typically, epidemiological studies for measuring pathogen fitness compute a single selection coefficient for pathogen variants that encompasses the entire lifecycle. However, evolution towards greater fitness within hosts can potentially compromise transmissibility. To address this, we develop a population-genetic model and software for the inference of multi-level selection coefficients. Our approach builds on population-genetic models previously applied to macro-evolution, whereby speciation and extinction rates are state-dependent. We relate this state dependence to within- and between-host selection. Evolution of beneficial phenotypes within hosts may occur rapidly (higher ‘speciation’), but reduce rates of dispersal between hosts (higher ‘extinction’). This new coalescent-based method improves on existing multi-state speciation and extinction models by accounting for highly nonlinear population dynamics and variable sampling through time, making it suitable for epidemiology. We characterize the accuracy of this method using birth–death and coalescent simulations before analysing a real dataset of HIV-1 genetic sequences. We quantified the multi-level relative fitness underlying HIV-1 ‘co-receptor switching’. In accordance with previous observations, we estimate that the relative transmissibility of X4-tropic virus is 9%, while within-host switching to X4-tropism is approximately 7-fold higher than switching to R5-tropism, indicating a strong within-host and between-host trade-off of co-receptor usage.

Background: Injection drug use (IDU) remains a major mode of HIV transmission in the United States, with rising use putting more people who inject drugs (PWID) at risk for HIV. Understanding HIV transmission clusters in this population can inform public health outreach strategies. We examined factors associated with linkage to HIV clusters among PWID in North Carolina (NC). Setting: NC maintains a statewide molecular HIV surveillance system, enabling identification of transmission clusters across diverse urban and rural communities. Methods: We analyzed data from people ≥13 years old, diagnosed with HIV, residing in NC, with ≥1 HIV sequence collected between April 2010 and March 2023. Genetic clusters were defined as ≥2 sequences with a pairwise genetic distance <1.5%. Poisson regression was used to estimate prevalence ratios for factors associated with cluster linkage. A logistic generalized additive model was then used to analyze the trend in cluster linkage among PWID over time. Results: Among 21,245 people with a sequence, 1,526 clusters were identified; 284 (19%) included ≥1 PWID (size range 2–218). Cluster linkage was associated with non-Hispanic White race/ethnicity, recent sequence collection, and factors associated with early HIV infection (high CD4 T-cell count, high viral load, and acute infection). The proportion of PWID linked to a cluster increased by year of HIV sequence collection. Conclusion: Early HIV testing and care are important for reducing transmission among PWID. Increased linkage of PWID to clusters likely reflects expansion of sequencing, which supports use of molecular clustering to identify groups needing additional outreach.

We investigate the predictability of evolution in terms of the phylogenetic placement of new lineages. This leads us to develop a class of coalescent models that relax neutrality by allowing the rate of coalescence to vary as a continuous heritable trait. In this setting, each lineage has a relative propensity to coalesce, with coalescent odds ratios defined from the product of pairwise propensities. Estimated coalescent odds provide a statistic that captures variation in lineage growth and are informative about the strength of natural selection acting on individual lineages. A number of practical statistical methods are then developed: techniques to adjust for biased and nonuniform sampling; procedures to automatically calibrate hyperparameters governing the evolution of coalescent propensity; and methods for clustering phylogenies into sets that delineate clades according to coalescent propensity. Simulations show sensitivity of these methods to detecting small selective effects acting on rare variants and strong robustness to imbalanced sampling. We demonstrate these methods using two datasets reflecting microbial populations evolving under strong selection. First, we examine a large set of Neisseria gonorrhoeae genomes and show that lineages with high coalescent odds feature a unique antibiotic resistance pattern, which presaged its subsequent expansion. We then reanalyze SARS-CoV-2 data in combination with independent estimates of reproduction numbers corresponding to major variants of concern 2020–2023. This indicates that coalescent odds can function as an excellent tree-based proxy for relative fitness of major SARS-CoV-2 lineages.

Motivation: How population structure can shape genetic diversity is a longstanding problem in population genetics. While the use of geographic locations, when available, can help answer some of these questions, it is still difficult to determine population structure when such metadata are not available or when the potential population structure is not easily observed. Here, we present an updated version of treestructure, an R package that implements a statistical test based on coalescent theory to detect unobserved population structure in a time-scaled phylogenetic tree. Availability: treestructure is available at CRAN at https://cloud.r-project.org/web/packages/treestructure/ and at https://emvolz-phylodynamics.github.io/treestructure/.

Microbial population genetic studies often involve the use of a dated phylogeny to show how the genomes are related over a relevant timescale. Many tools have recently been developed to date the nodes of a standard phylogeny, but all make underlying assumptions that may not be realistic for a given dataset, making the results potentially unreliable. Model comparison is sometimes used to remedy this issue, whereby inference under several models is compared to establish which result can be trusted. Although such comparison is clearly useful to assess the relative merits of several inference attempts, here instead we focus on the problem of evaluating how good an inference is in absolute terms, without comparison. We consider several approaches for diagnosing potential issues in a reconstructed dated phylogeny, including outlier detection, posterior predictive checking and residual analysis. These methods are well-established diagnostics tools in other areas of statistics, but here we show how they can be applied to the specific inference of dated phylogenies. We illustrate their use on many simulated datasets, with inference being performed either from the correct model to quantify the specificity or from an incorrect model to quantify the sensitivity of the diagnostics methods. We also applied the methods to three real-life datasets to showcase the range of issues that they can detect. We have implemented the methods in a new R package entitled DiagnoDating. We advocate the use of these diagnostics tools for all microbial population genetic studies that involve the reconstruction of a dated phylogeny.

We investigate the predictability of evolution in terms of the phylogenetic placement of new lineages. This leads us to develop a class of coalescent models that relax neutrality by allowing the rate of coalescence to vary as a continuous heritable trait. In this setting, each lineage has a relative propensity to coalesce, with coalescent odds ratios defined from the product of pairwise propensities. Estimated coalescent odds provide a statistic that captures variation in lineage growth and are informative about the strength of natural selection acting on individual lineages. A number of practical statistical methods are then developed: techniques to adjust for biased and non-uniform sampling; procedures to automatically calibrate hyperparameters governing the evolution of coalescent propensity; and, methods for clustering phylogenies into sets that delineate clades according to coalescent propensity. Simulations show sensitivity of these methods to detecting small selective effects acting on rare variants, and strong robustness to imbalanced sampling. We demonstrate these methods using two datasets reflecting microbial populations evolving under strong selection. First, we examine a large set of Neisseria gonorrhoeae genomes, and show that lineages with high coalescent odds feature a unique antibiotic resistance pattern which presaged its subsequent expansion. We then re-analyse SARS-CoV-2 data in combination with independent estimates of reproduction numbers corresponding to major variants of concern 2020-2023. This indicates that coalescent odds can function as an excellent tree-based proxy for relative fitness of major SARS-CoV-2 lineages.

Microbial population genetic studies often involve the use of a dated phylogeny to show how the genomes are related over a relevant timescale. Many tools have recently been developed to date the nodes of a standard phylogeny, but all make underlying assumptions that may not be realistic for a given dataset, making the results potentially unreliable. Model comparison is sometimes used to remedy this issue, whereby inference under several models is compared to establish which result can be trusted. Although such comparison is clearly useful to assess the relative merits of several inference attempts, here instead we focus on the problem of evaluating how good an inference is in absolute terms, without comparison. We consider several approaches for diagnosing potential issues in a reconstructed dated phylogeny, including outlier detection, posterior predictive checking and residual analysis. These methods are well-established diagnostics tools in other areas of statistics, but here we show how they can be applied to the specific inference of dated phylogenies. We illustrate their use on many simulated datasets, with inference being performed either from the correct model to quantify the specificity or from an incorrect model to quantify the sensitivity of the diagnostics methods. We also applied the methods to three real-life datasets to showcase the range of issues that they can detect. We advocate the use of these diagnostics tools for all microbial population genetic studies that involve the reconstruction of a dated phylogeny.

Background: HIV molecular cluster detection in the United States relies on HIV sequences obtained from drug resistance testing during clinical care (“routine care sequences”). This approach misses people who are not linked to care or who receive care but have uncollected or unreported sequences. Methods: We collected “HIV test sequences” from remnant serum samples of people testing newly positive from 2018 through 2021 by a large public health laboratory in North Carolina. We incorporated the HIV test sequences into a statewide molecular cluster analysis and assessed impact on “active cluster” detection (≥5 members newly diagnosed). We described data gaps filled by HIV test sequences, comparing (1) the extent of care sequence missingness due to gaps in care linkage vs sequence collection or reporting and (2) the characteristics of people with an HIV test sequence who had a care sequence, care but no care sequence, or no evidence of care. Results: Of 19 770 people included in the cluster analysis, 847 had an HIV test sequence, one-third of whom had no routine care sequence. We identified 13 additional active clusters (a 33% relative increase) and 40 larger active clusters after incorporating HIV test sequences. Most people with an HIV test sequence but no care sequence (78%) had another care indicator, suggesting sequence undercollection or underreporting, but a fifth (22%) had no evidence of care. Conclusions: Higher sequence coverage can improve cluster detection. While increased routine care sequence collection and reporting could fill many data gaps, sequencing remnant HIV test samples could include people without care linkage.

The fitness of a pathogen can be characterised at multiple levels corresponding to different stages of its life cycle, including replication within target cells (replicative capacity), immune evasion, and the capacity to infect new hosts (transmissibility). Pathogens evolve towards greater fitness within hosts, which can potentially compromise transmissibility. Typically, epidemiological studies for measuring pathogen fitness compute a single selection coefficient for pathogen variants that encompasses all stages of the pathogen life cycle. To address this problem, we develop a population-genetic model and software for the inference of multi-level selection coefficients which represent relative within-host fitness and transmissibility. Our approach builds on population genetic models previously applied to macro-evolution, whereby speciation and extinction rates are state-dependent. We relate the state-dependence in speciation and extinction rates to competing within- and between-host selection. Evolution of beneficial phenotypes within hosts may occur rapidly (higher ‘speciation’), but reduce rates of dispersal between hosts (higher ‘extinction’). This new coalescent-based method also accounts for highly nonlinear-population dynamics and variable sampling through time making it suitable for epidemiological studies. We derive the equilibrium proportion of a variant with multi-level selection under mutation-selection balance and show how this can be used in combination with the coalescent to greatly increase accuracy and precision of estimated selection coefficients. We characterise the accuracy of this method using a variety of birth-death and coalescent simulations before analysing a real dataset of HIV-1 genetic sequences from a large population-based sample of 4056 people living with HIV in Zambia. We quantified the multi-level relative fitness underlying HIV-1 ‘co-receptor switching’. In accordance with previous observations, we estimate that the relative transmissibility of X4-tropic virus is 9%, while within-host switching to X4-tropism is approximately seven-fold higher than switching to R5-tropism, indicating a strong within-host and between-host tradeoff of co-receptor usage. Our methods are available in an open-source R package musseco, providing a general-purpose tool for inferring multi-level selection coefficients for epidemiological sequence data. Author Summary The fitness of a pathogen can be characterised at each stage of its life cycle. However, epidemiological studies measuring pathogen fitness typically estimate a single value encompassing the entire life cycle. Pathogens evolve towards greater fitness within hosts, but this can compromise the capacity to infect new hosts. Our new method quantifies relative fitness at both levels. This method accounts for rapidly changing pathogen populations and variable sampling through time making it suitable for epidemiological studies. We tested our method using simulations and demonstrate its application by analysing real HIV-1 genetic sequences from people living with HIV in Zambia. This method sheds new light on long-standing questions about the relative fitness of HIV-1 variants which utilise different host-cell co-receptors. Using our new method, we estimate that the capacity of X4 variants to infect new people is just 9% of that of R5 variants, while within-host switching from R5 to X4 usage is seven times greater than the reverse switch. This indicates a strong within-host and between-host trade-off in co-receptor usage.

The evolution of HIV-1 virulence has significant implications for epidemic control. Recent phylogenomic analyses identified low-prevalence HIV-1 variants exhibiting significant differences in disease progression. We analysed 40 888 partial HIV-1 pol sequences from the UK HIV Drug Resistance Database (UKRDB) across subtypes B, C, A1, and CRF02AG. We identified phylotypes with putative differences in transmission/phylogenetic patterns and assessed their virulence trends using pretreatment viral loads, CD4 cell counts, and four statistical methods. We classified three subtype B phylotypes—PT.B.40.UK, PT.B.69.UK, and PT.B.133.UK —as variants of interest (VOIs) due to significantly higher viral loads and/or accelerated CD4 decline. PT.B.40.UK and PT.B.69.UK exhibited higher viral loads, 4.93 log10 copies/ml (95% CI: 4.73–5.13) and 4.87 (4.65–5.10), representing 0.30–0.36 log10 copies/ml higher than the reference group (4.57; 4.55–4.59). Despite uncertainties in baseline CD4 counts, all three VOIs reached the clinically relevant threshold of 350 CD4 cells/mm3 significantly faster than the reference group (3.5 years, 3.1–3.9 years): 2.3 years (1.0–5.1) for PT.B.40.UK, 2.0 years (10.8 months–4.4 years) for PT.B.69.UK, and 1.8 years (10.8 months–3.6 years) for PT.B.133.UK. These VOIs and their closest relatives have been circulating in the UK for decades with limited international spread and did not exhibit unusually rapid growth rates. Although these findings suggest a heritable high-virulence HIV-1 phenotype, we did not find evidence that convergent genetic polymorphisms or switches in coreceptor usage explained these differences. The small fraction of HIV-1 subtype B variants in the UK evolving towards higher virulence is unlikely to pose a public health concern, given the ongoing decline in new HIV diagnoses following the widespread adoption of pre-exposure prophylaxis and targeted prevention campaigns. However, this study—alongside the detection of the VB variant in the Netherlands—demonstrates that more virulent variants are not rare and can emerge independently in multiple countries. Consequently, HIV-1 genomic surveillance remains crucial to monitor HIV-1 virulence and mitigate its healthcare impact.