A Spatial Statistical Model for Landscape Genetics

TLDR

Landscape genetics seeks to link environmental features to population genetic structure, yet efficient methods for detecting genetic discontinuities are lacking. The article clarifies the conceptual framework for spatial modeling of genetic data. A Bayesian MCMC model treats sampled individuals as a spatial mixture of panmictic populations, using colored Voronoi tessellation to locate discontinuities, quantify spatial dependence, estimate the number of populations, assign individuals, and detect migrants, and is validated on simulated data. The method performs well on simulated datasets across a range of differentiation levels and successfully identifies population structure and migration in an 88‑individual wolverine sample.

Abstract

Landscape genetics is a new discipline that aims to provide information on how landscape and environmental features influence population genetic structure. The first key step of landscape genetics is the spatial detection and location of genetic discontinuities between populations. However, efficient methods for achieving this task are lacking. In this article, we first clarify what is conceptually involved in the spatial modeling of genetic data. Then we describe a Bayesian model implemented in a Markov chain Monte Carlo scheme that allows inference of the location of such genetic discontinuities from individual geo-referenced multilocus genotypes, without a priori knowledge on populational units and limits. In this method, the global set of sampled individuals is modeled as a spatial mixture of panmictic populations, and the spatial organization of populations is modeled through the colored Voronoi tessellation. In addition to spatially locating genetic discontinuities, the method quantifies the amount of spatial dependence in the data set, estimates the number of populations in the studied area, assigns individuals to their population of origin, and detects individual migrants between populations, while taking into account uncertainty on the location of sampled individuals. The performance of the method is evaluated through the analysis of simulated data sets. Results show good performances for standard data sets (e.g., 100 individuals genotyped at 10 loci with 10 alleles per locus), with high but also low levels of population differentiation (e.g., FST<0.05). The method is then applied to a set of 88 individuals of wolverines (Gulo gulo) sampled in the northwestern United States and genotyped at 10 microsatellites.