Classifying individuals among infra-specific taxa using microsatellite data and neural networks

Area of Science:

• Population Genetics
• Bioinformatics
• Machine Learning Applications

Background:

• Accurate individual assignment is crucial for population genetics studies.
• Traditional methods may have limitations in classifying individuals across diverse taxonomic levels.
• Microsatellite loci provide valuable genetic markers for population structure analysis.

Purpose of the Study:

• To evaluate the efficacy of neural networks for assigning individuals based on multilocus genotypes.
• To compare the performance of neural networks against discriminant analysis for individual assignment.
• To assess the accuracy of neural networks across different taxonomic levels in honeybee populations.

Main Methods:

• Utilized a dataset of 430 honeybees with 8 microsatellite loci.
• Applied two data transformation methods: simple coding and coding plus factorial correspondence analysis.
• Employed 'leave one out' and 'hold out' cross-validation procedures for prediction quality assessment.
• Compared neural network performance with discriminant analysis.

Main Results:

• Neural networks demonstrated superior performance in correctly classifying individuals compared to discriminant analysis.
• With simple coding and the 'hold out' procedure, classification accuracy reached 66.2% at the population level, 82.3% at the subspecies level, and 100% at the lineage level.
• The method proved effective across various taxonomical levels, including populations, subspecies, and evolutionary lineages.

Conclusions:

• Neural networks offer a powerful and accurate tool for individual assignment in population genetics.
• The findings suggest significant potential for applying neural networks to genetic data analysis.
• This approach can improve the understanding of population structure and evolutionary relationships.