Sampling bias and model choice in continuous phylogeography: Getting lost on a random walk

Abstract

Phylogeographic inference allows reconstruction of past geographical spread of pathogens or living organisms by integrating genetic and geographic data. A popular model in continuous phylogeography—with location data provided in the form of latitude and longitude coordinates—describes spread as a Brownian motion (Brownian Motion Phylogeography, BMP) in continuous space and time, akin to similar models of continuous trait evolution. Here, we show that reconstructions using this model can be strongly affected by sampling biases, such as the lack of sampling from certain areas. As an attempt to reduce the effects of sampling bias on BMP, we consider the addition of sequence-free samples from under-sampled areas. While this approach alleviates the effects of sampling bias, in most scenarios this will not be a viable option due to the need for prior knowledge of an outbreak’s spatial distribution. We therefore consider an alternative model, the spatial Λ-Fleming-Viot process (ΛFV), which has recently gained popularity in population genetics. Despite the ΛFV’s robustness to sampling biases, we find that the different assumptions of the ΛFV and BMP models result in different applicabilities, with the ΛFV being more appropriate for scenarios of endemic spread, and BMP being more appropriate for recent outbreaks or colonizations.

Dispersal dynamics of SARS-CoV-2 lineages during the first epidemic wave in New York City

“…With this procedure, we thus aimed (i) to construct and analyze data subsets that are amenable for joint Bayesian phylogeographic inference, and (ii) to explicitly mitigate sampling bias by subsampling NYC boroughs according to their relative importance during the epidemiological phase under investigation. Failing to correct for sampling bias can potentially lead to artifactual outcomes in the phylogeographic reconstructions [ 10 , 11 ]. For each of the ten subsets, we performed both a continuous [ 9 ] and a discrete [ 8 ] phylogeographic inference.…”

Exploiting genomic surveillance to map the spatio-temporal dispersal of SARS-CoV-2 spike mutations in Belgium across 2020

“…Another limitation in phylogeographic analyses is the possible effect of sampling bias. Hence, any conclusions have to be carefully interpreted and situated against the sampling pattern and sampling effort, although recent work in the field have tried to mitigate the effects of sampling bias 29 , 30 .…”

On the Use of Phylogeographic Inference to Infer the Dispersal History of Rabies Virus: A Review Study

“…Sampling bias can also be an issue in continuous phylogeographic inference when (a proportion of) samples are missing from certain locations, for example. Kalkauskas et al [ 80 ] have recently proposed to add sequence-free samples from undersampled areas to mitigate this issue, but acknowledge that it may be difficult to assess how many sequences to add from those locations. As an alternative to still perform such inference, the authors propose to use an alternative spatial -Fleming-Viot ( FV) process, which is unexplored for use in phylogeographic inference.…”

“…As an alternative to still perform such inference, the authors propose to use an alternative spatial -Fleming-Viot ( FV) process, which is unexplored for use in phylogeographic inference. While this model currently lacks an implementation in popular software applications, Kalkauskas et al [ 80 ] show it to be appropriate to model viral diffusion in cases of endemic spread, with the classic continuous phylogeographic model [ 13 ] being more appropriate for recent outbreaks.…”

“…For discrete phylogeographic analysis [ 12 ], typically one sampling location is assigned to a sequence, although in the case of uncertainty multiple locations can be considered with equal probability (using ambiguity codes), as well as using specific probabilities based on different sources of information [ 78 ]. In the atypical case of having a known RABV infection with available location data but missing genomic data, the use of “ghost” or “sequence-free” samples can be considered [ 79 , 80 ].…”