Understanding gene flow can help biodiversity to mitigate habitat changes by contributing to inform and design protected areas. The brown trout, Salmo trutta, displays a multitude of life-history strategies and represents an ideal model for applications in conservation genetics. Using a panel of 185-single nucleotide polymorphism markers, the present study aimed to explore the population structure of the brown trout and in the English Channel. The genotypes of 2,729 individual trout from 88 rivers were obtained across England and France. Population structure revealed the presence of genetic clusters following an east/west gradient. The maximum threshold distance between genetic distance and geographic distance was 344 km. The measure appeared relative to the studied spatial environment and reflected Salmo trutta capacity to achieve long migration distances. A machine-learning framework derived from a gradient forest analysis was used to generate a resistance surface using changes in allelic frequencies and environmental predicators. The resulting surface identified areas limiting gene flow. On the British coast, a genetic break was observed along the Jurassic coast, whereas the Cotentin peninsula acted as a physical barrier among French coastal populations. Salmo trutta populations appeared to be differently affected by environmental factors reflecting demes preference to specific breeding ground. Using our resistance map, the distance of maximum correlation using cost distance were computed allowing the pruning of our genetic graph. The resulting least cost path connections were mapped to reveal the main dispersal routes. Finally, a prioritization analysis using connectivity surface was implemented to design potential protected areas.

Understanding landscape connectivity has become a global priority for mitigating the impact of landscape fragmentation on biodiversity. Link-based methods traditionally rely on relating pairwise genetic distance between individuals or demes to their landscape distance (e.g., geographic distance, cost distance). In this study, we present an alternative to conventional statistical approaches to refine cost surfaces by adapting the Gradient Forest (GF) approach to produce a resistance surface. Used in community ecology, GF is an extension of random forest (RF), and has been implemented in genomic studies to model species genetic offset under future climatic scenarios. By design, this adapted method, resGF, has the ability to handle multiple environmental predicators and is not subjected to traditional assumptions of linear models such as independence, normality and linearity. Using genetic simulations, resGF performance was compared to other published methods. In univariate scenarios, resGF was able to distinguish the true surface contributing to genetic diversity among competing surfaces better than the compared methods. In multivariate scenarios, the GF approach performed similarly to the other RF-based approach using least-cost transect analysis (LCTA). Additionally, two worked examples are provided using two previously published datasets. This machine learning algorithm has the potential to improve our understanding of landscape connectivity and can inform long-term biodiversity conservation strategies.

Understanding landscape connectivity has become a global priority for mitigating the impact of landscape fragmentation on biodiversity. Link-based methods traditionally rely on relating pairwise genetic distance between individuals or demes to their landscape distance (e.g., geographic distance, cost distance). In this study, we present an alternative to conventional statistical approaches to refine cost surfaces by adapting the Gradient Forest (GF) approach to produce a resistance surface based on multiple environmental variables. Used in community ecology, Gradient Forest is an extension of random forest (RF), and has been implemented in genomic studies to model species genetic offset under future climatic scenarios. By design, this adapted method, resGF, has the ability to handle multiple environmental predictors and is not subjected to traditional assumptions of linear models such as independence, normality and linearity. Using genetic simulations, our RF-based approach was compared to other published methods using univariate and multivariate scenarios. Additionally, two worked examples are provided using two previously published datasets. This machine learning algorithm has the potential improve our understanding of landscape connectivity and can inform long-term biodiversity conservation strategies.