Bayesian optimization for demographic inference

Area of Science:

• Population Genetics
• Computational Biology
• Evolutionary Biology

Background:

• Demographic history inference is crucial in population genetics.
• Evaluating log-likelihoods for population genetics models is computationally intensive, especially with many populations.
• Existing genetic algorithm methods struggle with more than three populations.

Purpose of the Study:

• To develop a novel optimization pipeline for demographic inference.
• To address the computational challenges of evaluating time-consuming log-likelihoods.
• To improve inference accuracy for complex models involving multiple populations.

Main Methods:

• Utilized Bayesian optimization, a technique for optimizing expensive black-box functions.
• Developed a new optimization pipeline tailored for computationally demanding log-likelihood evaluations.
• Integrated the pipeline with the 'moments' tool for log-likelihood calculations.

Main Results:

• The new pipeline demonstrates superior performance compared to genetic algorithms.
• Superiority is evident in limited time budget scenarios with four and five populations.
• The method effectively handles complex demographic models with increased population numbers.

Conclusions:

• Bayesian optimization offers a powerful alternative for demographic inference.
• The developed pipeline enhances efficiency and accuracy for complex population genetic models.
• This approach is particularly beneficial when computational resources are limited.