Detecting Positive Selection by Modeling Structure Within Images of Genetic Variation.

Area of Science:

• Population genomics
• Bioinformatics
• Machine learning

Background:

• Accurately identifying natural selection from genomic data is a key challenge in population genomics.
• Dense whole-genome datasets allow for detailed analysis of genetic variation.
• Supervised machine learning methods can identify traces of natural selection but may lose spatial resolution.

Purpose of the Study:

• To introduce SKINET, a novel machine learning framework for detecting and characterizing positive natural selection.
• To address the loss of spatial resolution in existing methods like convolutional neural networks.
• To apply SKINET to human genome variation data for identifying adaptive genes.

Main Methods:

• Developed SKINET, integrating a trend filter kernel into a support vector machine framework.
• Trend filtering models feature autocovariation, preserving spatial relationships without architectural extensions.
• Applied SKINET to detect positive selection and estimate adaptive parameters.

Main Results:

• SKINET effectively distinguishes regions under positive natural selection from neutral regions.
• The method functions in a regression framework to estimate associated adaptive parameters.
• Application to human data identified known adaptive genes and novel targets like FAM177A1, linked to cancer.

Conclusions:

• SKINET offers a powerful and spatially accurate approach for detecting natural selection in genomic data.
• The method successfully identifies adaptive candidate genes, including novel cancer-related targets.
• SKINET advances the analysis of population genomics and evolutionary adaptation.