Pitfalls and Remedies for Cross Validation with Multi-trait Genomic Prediction Methods

Area of Science:

• Quantitative genetics
• Genomic prediction
• Plant and animal breeding

Background:

• Multi-trait genomic prediction models can enhance prediction accuracy and selection gain.
• However, these models are complex and can overfit, potentially reducing accuracy compared to single-trait models.
• Cross-validation is standard for model selection but can be biased in multi-trait scenarios.

Purpose of the Study:

• To demonstrate the bias in naive cross-validation for multi-trait genomic prediction.
• To propose and evaluate novel cross-validation strategies for accurate model assessment.
• To guide the effective use of multi-trait information in genomic selection.

Main Methods:

• Simulations were used to illustrate the bias of naive cross-validation.
• Three solutions were proposed: a parametric selection index approach, a semi-parametric accuracy correction, and a non-parametric method (CV2*).
• CV2* involves validating predictions against genetically related individuals' focal trait measurements.

Main Results:

• Naive cross-validation can lead to severely biased estimates of prediction accuracy in multi-trait models.
• The proposed methods offer partial corrections to improve the reliability of cross-validation.
• CV2* provides a robust alternative for validating multi-trait genomic predictions.

Conclusions:

• Appropriate cross-validation is crucial for reliably determining the utility of multi-trait genomic prediction.
• The proposed methods can help mitigate bias and improve model selection in breeding programs.
• Accurate assessment of multi-trait models is essential for leveraging high-throughput phenotyping data.