Deciphering signatures of natural selection via deep learning

Area of Science:

• Genomics
• Evolutionary Biology
• Bioinformatics

Background:

• Identifying genomic regions under natural selection is key to understanding local adaptation.
• Detecting loci influenced by complex, spatially varying selection remains a significant challenge in population genetics.

Purpose of the Study:

• To introduce DeepGenomeScan, a novel deep learning framework for detecting signatures of spatially varying selection.
• To evaluate DeepGenomeScan's performance against established genome scan methods.

Main Methods:

• Development and application of a deep learning-based framework, DeepGenomeScan.
• Comparative analysis with principal component analysis (PCA), redundancy analysis (RA), SPA, iHS, Fst, and Bayenv.

Main Results:

• DeepGenomeScan outperformed PCA- and RA-based methods in identifying loci under complex spatial selection.
• Statistical power increased by up to 47.25% under nonlinear environmental selection.
• Identified known and novel clinically important genes in a European dataset, surpassing traditional methods.

Conclusions:

• DeepGenomeScan offers a powerful new approach for detecting spatially varying selection.
• The framework enhances the discovery of genes underlying local adaptation, including those with clinical relevance.