Infectious disease phylodynamics with occurrence data

Abstract

Phylodynamic models use pathogen genome sequence data to infer epidemiological dynamics. With the increasing genomic surveillance of pathogens, especially during the SARS‐CoV‐2 pandemic, new practical questions about their use are emerging. One such question focuses on the inclusion of un‐sequenced case occurrence data alongside sequenced data to improve phylodynamic analyses. This approach can be particularly valuable if sequencing efforts vary over time. Using simulations, we demonstrate that birth–death phylodynamic models can employ occurrence data to eliminate bias in estimates of the basic reproductive number due to misspecification of the sampling process. In contrast, the coalescent exponential model is robust to such sampling biases, but in the absence of a sampling model it cannot exploit occurrence data. Subsequent analysis of the SARS‐CoV‐2 epidemic in the northwest USA supports these results. We conclude that occurrence data are a valuable source of information in combination with birth–death models. These data should be used to bolster phylodynamic analyses of infectious diseases and other rapidly spreading species in the future.

Decoding the fundamental drivers of phylodynamic inference

“…The results of our simulation study add clarity to previous work showing that sampling times contribute substantially to phylodynamic inference under the birth death (Volz and Frost, 2014;Featherstone et al, 2021). We demonstrate that lower sequence diversity often precludes sequence data from a comparable effect.…”

Phylogenetic and phylodynamic approaches to understanding and combating the early SARS-CoV-2 pandemic

“…The impact of low and uneven sampling intensity on SARS-CoV-2 phylogeography has been recognized 25 , 99 , and is commonly addressed through downsampling or a bespoke sampling regimen 99 . Some phylodynamic models allow for explicit modelling of sampling bias, and this has enhanced several SARS-CoV-2 studies 12 , 25 , 100 . In some cases, a single uniquely shared variant may be sufficient to determine the origin of a transmission chain 60 , 101 , and in studies in which one or several regions or periods were poorly sampled, unsampled individuals have been modelled and added to analyses 102 , or unobserved ancestral locations jointly inferred using phylogenies 103 .…”

“…Uneven sampling through time can be addressed by adding an explicit sampling model to phylodynamic inference. One current solution uses structured (epoch-based) models to condition on the rate of genomic sampling relative to all PCR-confirmed SARS-CoV-2 cases, and reportedly improves molecular clock accuracy 100 . Methods that can accommodate changing rate of sequencing through time have been developed, for example, the coalescent-based Bayesian Epoch Skyline Plot (ESP) 109 (see also Box 1 ), an approach analogous to the classic BD-skyline 110 .…”

“…Methods that can accommodate changing rate of sequencing through time have been developed, for example, the coalescent-based Bayesian Epoch Skyline Plot (ESP) 109 (see also Box 1 ), an approach analogous to the classic BD-skyline 110 . An alternative is to model sampling while linking sample location to regional variations in sampling effort; this has improved estimation of population size history for at least some data sets 111 , and the BD-skyline with variable sampling rate has also been applied to SARS-CoV-2 (refs 100 , 112 ). The relationship between genetic variation and transmission patterns is one of interdependence, and therefore combining phylodynamic estimation with epidemiological data should generate stronger inferences.…”

Statistical Challenges in Tracking the Evolution of SARS-CoV-2

“…In Frost et al (2015), the authors stated the following challenges: (1) modeling of more complex evolutionary processes such as recombination, selection, within-host evolution, population structure, and stochastic population dynamics; (2) modeling of more complex sampling scenarios, (3) joint modeling of phenotypic and genetic data, and (4) computation. We have subsequently seen advances in solving some of these challenges, such as modeling of recombination (Müller, Kistler and Bedford, 2021) and stochastic population dynamics Volz and Siveroni, 2018), incorporation of more complex sampling scenarios (Karcher et al, 2016(Karcher et al, , 2020Parag, du Plessis and Pybus, 2020;Cappello and Palacios, 2021), and joint modeling of epidemiological and genetic data (Li, Grassly and Fraser, 2017;Tang et al, 2019;Zarebski et al, 2021;Featherstone et al, 2021). However, even in the simplest evolutionary model, inference involves integration over the high dimensional space of phylogenies.…”