Dispersal inference from population genetic variation using a convolutional neural network.

Area of Science:

• Ecology
• Population Genetics
• Bioinformatics

Background:

• Biological dispersal patterns influence genetic variation across landscapes.
• Estimating dispersal distance is crucial for understanding population dynamics and conservation.

Purpose of the Study:

• To introduce disperseNN, a novel deep learning tool for estimating mean per-generation dispersal distance from genotype data.
• To evaluate the performance of disperseNN against existing methods and assess its robustness to nuisance parameters.

Main Methods:

• Utilized a convolutional neural network trained on geographically distributed genotype data.
• Input data included single-nucleotide polymorphism (SNP) data and spatial sampling scale.
• Evaluated performance using extensive simulations and real-world data from 12 species.

Main Results:

• DisperseNN demonstrated competitive or superior performance compared to state-of-the-art methods, especially with small sample sizes.
• The method effectively estimated dispersal distance even when population density, demographic history, and habitat size were unknown.
• DisperseNN provided larger dispersal estimates than mark-recapture studies, potentially due to broader geographic scope.

Conclusions:

• DisperseNN offers a powerful, data-efficient tool for inferring dispersal distances in various species, complementing traditional methods.
• The method's independence from geographic coordinates and local population density makes it broadly applicable, particularly for non-model organisms.
• This genetic approach enhances our understanding of biological dispersal and landscape connectivity.