Genetic prediction with ARG-powered linear algebra

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Ancestral recombination graphs (ARGs) model genetic relatedness, drift, recombination, and mutation.
• Efficient ARG storage is crucial for processing large genomic datasets and fitting linear mixed models.

Purpose of the Study:

• To develop efficient algorithms for variance component estimation and genetic value prediction using ARGs.
• To apply these methods to biobank-scale phenotype and genome datasets.

Main Methods:

• A generative model for complex traits with additive effects on an ARG was described.
• Algorithms leveraging the succinct tree sequence representation of ARGs and randomized linear algebra were developed.
• Restricted maximum likelihood (REML) was used for variance component estimation.

Main Results:

• Algorithms demonstrated nearly linear runtime scaling with sample size.
• REML outperformed the Haseman-Elston method for variance component estimation.
• Inferred ARGs yielded variance component estimates and genetic predictions comparable to true ARGs in simulations.

Conclusions:

• The developed algorithms provide efficient solutions for quantitative genetic analysis on biobank-scale data using ARGs.
• Variance component estimates can be interpreted as mutational and additive genetic variance.
• The Python package tslmm implements these algorithms, utilizing the tskit library.