Detecting introgression from phylogenetic invariant site patterns using machine learning.

Area of Science:

• Evolutionary genetics
• Phylogenomics
• Bioinformatics

Background:

• Detecting historical introgression from genomic data is crucial in evolutionary genetics.
• Current methods like network inference and admixture inference have limitations in computational scalability and sample size.
• Existing admixture inference methods (e.g., ABBA-BABA tests) are limited to small sample quartets.

Purpose of the Study:

• To develop a novel computational method for detecting historical introgression.
• To overcome the limitations of existing phylogenetic inference methods.
• To analyze genomic data more comprehensively by considering all quartet information simultaneously.

Main Methods:

• Developed simcat, a machine learning model utilizing neural networks.
• Trained the model on coalescent simulations to identify introgression patterns from single-nucleotide polymorphism (SNP) frequencies.
• Evaluated simcat's performance on simulated data and an empirical dataset of oak trees (Quercus ser. Virentes).

Main Results:

• Simcat accurately classifies introgression events in simulations.
• The method demonstrates sensitivity to variations in species tree parameters.
• Successfully applied simcat to a real-world dataset of oak trees, showcasing its practical utility.

Conclusions:

• Introduced a novel machine learning approach (simcat) for historical introgression detection.
• Expanded phylogenetic invariants-based methods to a larger phylogenetic context beyond quartets.
• Paved the way for leveraging machine learning in large-scale phylogenetic analyses.