2.3 | Assessing the invasion history
The global invasion history of R. flavipes was inferred through
ABC analyses by comparing support for different invasion scenarios. The
scenarios varied according to the origin(s) of introduced populations,
the founding population size, the bottleneck duration and the admixture
rate if multiple sources were detected. To reduce computational effort,
model selection and parameter estimation were performed using the
recently developed random forests (RF) machine learning method (ABC RF)
available in the abcrf R package
(Pudlo et al. 2015,
Raynal et al. 2018). This method requires
a reduced number of simulated datasets while providing robust posterior
estimates. A step-by-step approach (9 different steps divided into 4
parts; fully explained in Supporting Information 1) was used to infer
the different episodes of the invasion history of R. flavipes, as
this type of approach is commonly performed in ABC studies to reduce
computational effort (Fraimout et al.
2017, Javal et al. 2019,
Ryan et al. 2019). The introduced
populations of Germany, Uruguay and the Bahamas were not used in ABC
computations as they were represented by too few individuals. Briefly,
the first part estimated whether each introduced population
(i.e., France, Canada and Chile) arose from independent or
bridgehead introduction events (Part A). As this first part indicated
that the French population may have played a role in the introduction to
Canada and Chile, we first sought to decipher the source(s) of
introductions to France alone (Part B). Then, we attempted to identify
the sources of the Canada (Part C) and Chile (Part D) populations using
France as a potential source. For all scenarios tested, introduction
events were followed by a decrease in effective size varying between one
to 100 migrants for a duration of zero to 50 years. Posterior
distributions of preliminary simulated data sets were used to adjust the
range of other priors as wide as possible while retaining biological
meaning. For all scenarios in each step, at least 10,000 simulated
datasets including all summary statistics implemented in the DIYABC
software v.2.1.0 (Cornuet et al. 2014)
were generated from 2,000 randomly sampled SNPs. Priors were set uniform
for all model parameters and selected based on historical records. The
different scenarios tested within each step are provided in the
Supplementary Information.