Semi-supervised detection of natural selection with positive-unlabeled learning

Area of Science:

• Evolutionary genomics
• Machine learning
• Population genetics

Background:

• Identifying genomic regions under natural selection is crucial in evolutionary genomics.
• Current machine learning methods rely on simulated data and explicit labels, limiting their application to real-world genomes with mixed evolutionary forces.
• One-vs.-rest strategies are complex and struggle to model diverse evolutionary backgrounds.

Purpose of the Study:

• To develop a flexible machine learning framework for detecting adaptive events, specifically selective sweeps, in genomic data.
• To introduce a positive-unlabeled learning approach that bypasses the need for explicit negative sample modeling.
• To enable robust detection of positive selection in realistic genomic landscapes shaped by various evolutionary factors.

Main Methods:

• Introduced PULSe, a method employing positive-unlabeled learning for selective sweep detection.
• Trained PULSe using only labeled examples of selective sweeps and an unlabeled background dataset.
• Evaluated PULSe's performance across diverse demographic, adaptive, and confounding scenarios, including domain shift.

Main Results:

• PULSe demonstrated high performance and generalizability across various genomic contexts.
• The method successfully identified previously known selective sweep candidates in European and Bengali human genomes.
• PULSe effectively handles complex genomic data without assumptions about the background composition.

Conclusions:

• PULSe offers a powerful and versatile alternative for detecting adaptive genomic regions.
• The positive-unlabeled learning framework provides robustness in realistic genomic analyses.
• PULSe has the potential to generalize across diverse genomic landscapes and evolutionary scenarios.