Machine Learning and Genomics: Using a Novel Neural Network for Population Assignment of a Threatened Seabird, Leach’s Storm-Petrel (Hydrobates leucorhous)
Loading...
Authors
Lounder, Heather
Date
Type
Language
eng
Keyword
Machine Learning , Neural network , Population assignment , genomics , seabird , leachs storm-petrel
Alternative Title
Abstract
Determining the population of origin for individuals, referred to as population assignment, can help researchers better understand and quantify threats facing different species. Population assignment has many broad applications, however previous methods have largely relied on genetic differences among populations to identify source populations for samples of unknown origin. While this may be suitable for highly structured species, for migratory species with weak population genetic differentiation, population assignment is much more challenging. Leach’s Storm-petrels (Hydrobates leucorhous) are small pelagic seabirds that breed in large colonies throughout the North Atlantic. They are currently threatened with extinction due to significant population declines observed over the past 50 years. Threats facing this species are multifaceted and difficult to quantify, and the impact of mortality on different colonies is currently unknown. Previous attempts to perform assignment to source colonies in this species have failed to generate the fine-scale assignments needed to quantify the impacts of mortality at different colonies. In this thesis, I used double digest restriction site-associated DNA sequencing (ddRADseq) to provide 11356 genome-wide single nucleotide polymorphisms (SNPs) to investigate the genetic relationships among 10 colonies of Leach’s Storm-petrels and perform genetic population assignment using novel classifier neural network popfinder. I used assignment results to determine if mortality experienced away from the breeding colony was disproportionately impacting certain colonies based on colony size. I found no genetic structure among colonies, suggesting that Atlantic storm-petrel colonies are genetically connected. The program popfinder was able to determine the breeding colony of origin for birds with high accuracy and precision (>80%), a task previously not possible in this species. Mortality appears to be having colony specific impacts, with 8 of the sampled colonies having a significantly different number of birds assign than would be expected based on colony size. Sources of mortality for this species are likely widespread, though more data are needed to perform colonyspecific threat assessment and determine if the distribution of mortality presented here is also present on a larger scale. Nonetheless, the results presented here represent a promising new method for genetic population assignment in weakly structured species.