A Likelihood-Free Inference Framework for Population Genetic Data using Exchangeable Neural Networks.

Area of Science:

• Population genetics
• Computational biology
• Machine learning

Background:

• High-throughput DNA sequencing has increased interest in population-scale inference.
• Existing methods struggle with complex models and high-dimensional data.
• Likelihood-free methods often rely on hand-designed summary statistics, limiting accuracy.

Purpose of the Study:

• To develop a scalable, general-purpose inference technique for complex population genetic models.
• To address challenges of data exchangeability and intractable likelihood computations.
• To create a summary statistic-free, likelihood-free inference framework.

Main Methods:

• Developed an exchangeable neural network architecture.
• Utilized a simulation-based, likelihood-free inference approach.
• Applied the framework in a black-box manner to various simulation tasks.

Main Results:

• The proposed method performs summary statistic-free, likelihood-free inference.
• The framework demonstrates broad applicability across biological and non-biological tasks.
• Achieved superior performance on the recombination hotspot testing problem compared to state-of-the-art methods.

Conclusions:

• The exchangeable neural network offers a powerful new approach for population genetics inference.
• This method overcomes limitations of traditional likelihood-free techniques.
• The framework provides accurate and scalable inference for complex population-scale genomic data.