Deep Learning for Population Genetic Inference.

Area of Science:

• Population genetics
• Genomics
• Machine learning

Background:

• Computing likelihoods for complex population genetic models with genomic variation data is computationally challenging.
• Existing methods often struggle to jointly infer demographic history and natural selection.
• Pervasive natural selection can confound demographic analyses, particularly in species like Drosophila.

Purpose of the Study:

• To introduce a novel likelihood-free inference framework using deep learning for population genetic analysis.
• To demonstrate the effectiveness of deep learning in inferring population genetic parameters and learning informative data features.
• To apply the framework for jointly inferring natural selection and demographic history in Drosophila melanogaster.

Main Methods:

• Developed a deep learning framework utilizing multilayer neural networks.
• Trained neural networks to map summary statistics of genomic data to population genetic parameters.
• Applied the framework to 197 African Drosophila melanogaster genomes to infer demography and selection.

Main Results:

• The deep learning method successfully separated global demographic effects from local selection signals.
• Identified numerous genomic regions in Drosophila melanogaster that have undergone hard selective sweeps.
• Found evidence for soft sweeps and balancing selection, particularly near chromosome centromeres.
• Demographic inference suggests previously estimated bottlenecks in African Drosophila melanogaster populations were overly severe.

Conclusions:

• Deep learning provides a powerful and effective approach for likelihood-free inference in population genetics.
• The developed framework enables simultaneous inference of demography and selection, overcoming limitations of sequential analyses.
• The study provides new insights into the demographic history and selective pressures shaping African Drosophila melanogaster populations.