Priors in whole-genome regression: the bayesian alphabet returns

Area of Science:

• Genomics
• Quantitative Genetics
• Statistical Genetics

Background:

• Whole-genome prediction (WGP) is crucial for complex traits in breeding and genetics.
• Bayesian linear regression models, termed the 'Bayesian alphabet', address overparameterization (p > n) in genomic data.
• Prior distributions significantly influence WGP models when parameters exceed sample size.

Purpose of the Study:

• To investigate the impact of prior distributions on WGP models for complex traits.
• To evaluate the inferential value and predictive accuracy of Bayesian alphabet models under typical genomic data conditions (p > n).

Main Methods:

• Exploration of various Bayesian linear regression models (Bayesian alphabet) with different priors.
• Analysis of WGP models in scenarios where the number of parameters (p) exceeds the sample size (n).
• Assessment of parameter influence and predictive performance through cross-validation.

Main Results:

• Prior distributions remain influential in WGP models when p > n, as parameters are not solely determined by the data (likelihood identification).
• Bayesian learning is imperfect, leading to potential caution regarding genetic architecture claims.
• Cross-validation is essential for reliable predictions, as 'tuning knobs' can optimize performance.

Conclusions:

• Bayesian alphabet models are valuable for WGP of phenotypes.
• Their inferential value for genetic architecture is questionable when sample size is limited (n << p).
• Careful cross-validation is recommended for robust predictions and reliable interpretation.