Detecting adaptive introgression in human evolution using convolutional neural networks.

Area of Science:

• Evolutionary Genetics
• Genomics
• Bioinformatics

Background:

• Adaptive introgression, the transfer of beneficial alleles between species, is a significant evolutionary mechanism.
• Detecting adaptive introgression using genomic data is challenging due to the lack of explicit modeling frameworks for introgression and positive selection.
• Existing methods struggle to jointly analyze introgression dynamics and selection signatures.

Purpose of the Study:

• To develop a novel computational framework for identifying adaptive introgression using genomic sequence data.
• To leverage convolutional neural networks (CNNs) for their ability to model complex patterns without predefined analytical models.
• To accurately distinguish genomic regions under adaptive introgression from neutral evolution or selective sweeps.

Main Methods:

• Developed a convolutional neural network (CNN) architecture trained on simulated genotype matrices.
• Utilized data from donor, recipient, and non-introgressed populations to train the CNN.
• Tested the CNN's performance on both phased and unphased genomic data, including scenarios with heterosis.

Main Results:

• The CNN architecture achieved 95% accuracy in identifying adaptive introgression in simulated data.
• Performance remained robust even with unphased genomes and showed moderate decreases with heterosis.
• Successfully applied the trained CNNs to human genomic datasets to identify potential adaptive introgression events.

Conclusions:

• Convolutional neural networks provide a powerful and accurate tool for detecting adaptive introgression.
• This method offers a significant advancement for analyzing genomic sequence data to understand evolutionary processes.
• The approach has implications for studying the role of introgression in shaping the evolutionary history of various species, including humans.