Inference of Coalescence Times and Variant Ages Using Convolutional Neural Networks.

Area of Science:

• Population Genetics
• Genomics
• Machine Learning

Background:

• Accurate inference of time to the most recent common ancestor (TMRCA) and genomic variant ages is crucial for population genetics.
• Existing model-based approaches have limitations in certain scenarios.

Purpose of the Study:

• To develop a novel, accurate, and adaptable method for inferring pairwise TMRCAs and allele ages.
• To apply this method to large-scale genomic data for insights into demographic history and selection.

Main Methods:

• Developed CoalNN, a likelihood-free approach using convolutional neural networks trained via simulation.
• Utilized transfer learning to adapt the model to varying demographic parameters.
• Applied CoalNN to 2,504 samples from the 1,000 Genomes Project, analyzing ~80 million variants.

Main Results:

• CoalNN matched or surpassed existing model-based methods in accuracy for TMRCA and allele age prediction across simulations.
• Inferred variant ages across 26 populations revealed significant variation, reflecting demographic histories and negative selection.
• Generated genome-wide annotations of negative selection signatures, improving heritability analysis for complex traits.

Conclusions:

• Likelihood-free, simulation-trained models are effective for inferring gene genealogy properties in large genomic datasets.
• CoalNN provides valuable insights into population demographics and evolutionary processes.
• The developed annotations enhance the study of heritability and the impact of selection on complex traits.