Automatic inference of demographic parameters using generative adversarial networks.

Area of Science:

• Population genetics
• Computational biology
• Bioinformatics

Background:

• Simulated data are vital for population genetics research, aiding validation and inference.
• Current simulation software requires extensive parameter tuning and often fails to capture real genetic data properties.
• This limitation hinders the development and application of advanced analytical methods.

Purpose of the Study:

• To develop a novel method for estimating parameters in population genetic models.
• To create synthetic genetic data that accurately mirrors real-world data properties.
• To enable automatic adaptation to diverse population genetic datasets.

Main Methods:

• Developed a generative adversarial network (GAN) approach named pg-gan.
• pg-gan learns to generate realistic synthetic genetic data.
• Applied the method to an isolation-with-migration model and human data from the 1000 Genomes Project.

Main Results:

• Successfully recovered input parameters in a simulated isolation-with-migration model.
• Demonstrated the ability of pg-gan to generate synthetic data that recapitulates features of real human genetic data.
• Validated the method's effectiveness on empirical data.

Conclusions:

• pg-gan offers a powerful new approach for generating realistic population genetic data.
• This method enhances the accuracy of parameter estimation and the scope of analyses in population genetics.
• The approach has significant implications for machine learning applications in evolutionary biology.