Applications of random forest feature selection for fine-scale genetic population assignment

Area of Science:

• Population genetics
• Genomics
• Conservation science

Background:

• Accurate genetic population assignment is crucial for wildlife management and conservation.
• Effective assignment relies on informative genetic markers and sensitive analytical tests.

Purpose of the Study:

• To compare machine learning algorithms (random forest, regularized random forest, guided regularized random forest) with FST ranking for selecting single nucleotide polymorphisms (SNPs).
• To determine the minimum SNP panel size for achieving at least 90% self-assignment accuracy in Atlantic and Chinook salmon populations.

Main Methods:

• Applied random forest-based methods and FST ranking to SNP datasets for Atlantic salmon (Salmo salar) and Chinook salmon (Oncorhynchus tshawytscha).
• Determined minimum panel sizes for achieving 90% self-assignment accuracy for each species and method.
• Evaluated performance differences between machine learning-selected and FST-selected SNP panels.

Main Results:

• Random forest-based SNP panels achieved higher self-assignment accuracy compared to FST-selected panels.
• Performance gains were up to 7.8 and 11.2 percentage points for Atlantic and Chinook salmon, respectively.
• Achieved 90% self-assignment accuracy with 670 and 384 SNPs for Atlantic and Chinook salmon, respectively, surpassing FST-based methods.

Conclusions:

• Machine learning approaches are effective for selecting informative SNPs for population assignment.
• These methods enhance genetic assignment accuracy, benefiting wildlife management and conservation of exploited populations.
• The study demonstrates the utility of machine learning in marker selection for large genomic datasets.