Predicting geographic location from genetic variation with deep neural networks

Area of Science:

• Genetics
• Bioinformatics
• Population Genetics

Background:

• Genetic data often exhibits spatial autocorrelation, meaning organisms are more related to nearby individuals.
• This spatial genetic structure allows for the prediction of an organism's geographic origin based on its genetic makeup.

Purpose of the Study:

• To introduce Locator, a novel deep learning method for predicting sample geographic origin.
• To demonstrate Locator's superior speed and accuracy compared to existing methods.
• To enable the analysis of mosaic ancestry and geographic mixing in populations.

Main Methods:

• Developed a deep learning model named Locator.
• Simulated genetic data to test Locator's performance.
• Applied Locator to whole-genome sequence data from Plasmodium parasites, Anopheles mosquitoes, and human populations.

Main Results:

• Locator infers sample location to within 4.1 generations of dispersal in simulations.
• Locator runs at least an order of magnitude faster than previous model-based approaches.
• Median test errors were 16.9km (Plasmodium), 5.7km (Anopheles), and 85km (humans).

Conclusions:

• Locator provides a fast and accurate method for inferring geographic origins from genetic data.
• The method's efficiency allows for genome-wide analysis, revealing mosaic ancestry and population mixing.
• Locator has broad applications in studying the spatial distribution and spread of various organisms.