Cross-Validation Without Doing Cross-Validation in Genome-Enabled Prediction.

Area of Science:

• Quantitative genetics
• Genomic prediction
• Statistical modeling

Background:

• Cross-validation is crucial for accurate genome-enabled prediction of complex traits.
• Current methods often require extensive model retraining, increasing computational cost.
• Efficient validation strategies are needed for large-scale genomic datasets.

Purpose of the Study:

• To develop novel formulae for computing cross-validation predictions without retraining models.
• To enable efficient evaluation of various genome-assisted prediction methods.
• To reduce computational burden in genomic prediction studies.

Main Methods:

• Derived formulae for leave-one-out and leave-many-out cross-validation predictions.
• Applied importance sampling for Bayesian models.
• Tested on a wheat dataset for grain yield prediction using 1279 markers.

Main Results:

• Demonstrated proof of concept for the developed formulae on a wheat dataset.
• Evaluated predictive mean-squared error and impact of regularization parameters.
• Showcased the efficiency of the new approach for extensive cross-validation.

Conclusions:

• The developed formulae significantly expedite cross-validation in genome-assisted prediction.
• This approach is applicable to a wide range of linear models, including BLUP and Bayesian alphabet methods.
• Facilitates more extensive and computationally feasible cross-validation for quantitative trait prediction.