Neural posterior estimation for population genetics.

Area of Science:

• Population Genetics
• Computational Biology
• Bioinformatics

Background:

• Simulation-based inference is crucial in population genetics, especially when likelihood-based methods fail.
• Approximate Bayesian Computation (ABC) is widely used but computationally intensive and struggles with high-dimensional data.
• Supervised machine learning (ML) offers an alternative but typically lacks Bayesian uncertainty estimation.

Purpose of the Study:

• To introduce and evaluate Neural Posterior Estimation (NPE) as a novel inference method in population genetics.
• To compare NPE's accuracy and efficiency against existing methods like ABC.
• To demonstrate NPE's utility in demographic inference and provide a user-friendly workflow.

Main Methods:

• Implemented a neural network to estimate posterior distributions for population genetics models.
• Compared NPE with other inference techniques using raw genotypes and summary statistics.
• Applied NPE to demographic inference, including a case study on Drosophila melanogaster.

Main Results:

• NPE achieved high accuracy and efficiency in estimating posterior distributions.
• Learned posterior distributions were successfully generated using both raw genotypes and summary statistics.
• NPE proved effective for both simple and complex demographic inference scenarios.

Conclusions:

• NPE successfully integrates the advantages of ABC and supervised ML for population genetics inference.
• The method provides accurate and efficient posterior distribution estimation, outperforming traditional approaches in certain aspects.
• NPE is a promising tool for diverse population genetics applications, including demographic history analysis, with a provided workflow for broader adoption.