Comparisons of likelihood and machine learning methods of individual classification

Area of Science:

• Population Genetics
• Machine Learning
• Bioinformatics

Background:

• Machine learning (ML) classification methods are infrequently applied to population genetic data.
• Traditional methods for assigning individuals to populations (assignment tests) often use parametric likelihood estimations.
• Comparing ML techniques with established methods is crucial for advancing population genetic analyses.

Purpose of the Study:

• To evaluate and compare the accuracy of nonparametric machine learning classifiers against parametric likelihood estimators for population assignment.
• To assess classifier performance across simulated datasets with varying population differentiation (FST), loci number, and allelic diversity.
• To examine classifier performance using empirical lake trout (Salvelinus namaycush) genetic data.

Main Methods:

• Compared artificial neural networks, decision trees, and k-nearest neighbor clustering with likelihood estimation methods.
• Utilized simulated population genetic datasets with controlled parameters (FST, loci, alleles).
• Applied methods to empirical lake trout genetic data exhibiting comparable population differentiation.

Main Results:

• Artificial neural networks and likelihood estimators exhibited lower classification error rates than k-nearest neighbor and decision trees on simulated data.
• Artificial neural networks showed only marginal improvement (0-2.8% lower error) over likelihood methods for simulated data.
• ML classifiers relatively outperformed likelihood estimators on empirical data, indicating an ability to leverage population-specific genotypic patterns.

Conclusions:

• While ML classifiers show promise, particularly on empirical data, likelihood-based methods remain more accessible and reliable for population assignment.
• The complexity in developing and evaluating artificial neural networks makes likelihood methods preferable for routine population genetic studies.
• Further research into ML applications may enhance population genetic analyses, but current likelihood approaches offer a robust and practical solution.