Predicting genetic values: a kernel-based best linear unbiased prediction with genomic data.

Area of Science:

• Genomics
• Statistical Genetics
• Bioinformatics

Background:

• Genomic data offers valuable insights for modeling covariance structures, enhancing the accuracy of total genetic value (GV) predictions.
• Traditional geostatistical prediction methods operate in low-dimensional spaces, necessitating adaptation for high-dimensional genomic data.

Purpose of the Study:

• To adapt geostatistical kriging concepts for high-dimensional genomic prediction using single nucleotide polymorphism (SNP) vectors.
• To evaluate the performance of Universal Kriging (UK) and Simple Kriging (SK) with Matérn covariance functions against Genomic Best Linear Unbiased Prediction (GBLUP).

Main Methods:

• Application of Universal Kriging (UK) and Simple Kriging (SK) to model SNP vector covariance structures using Matérn functions.
• Comparison with Genomic Best Linear Unbiased Prediction (GBLUP) as a reference method.
• Whole-genome simulation study evaluating additive, additive-dominance, and epistatic gene-action models.

Main Results:

• Universal Kriging (UK) demonstrated superior predictive performance over GBLUP, particularly in scenarios involving dominance and epistatic gene effects.
• The study showed that VanRaden's (2008) genomic covariance structure is a special case of the proposed covariance function with a quadratic variogram.
• Theoretical proof established a link between GVs from different linear mixed models when covariance matrices exhibit a specific linear relationship.

Conclusions:

• Kriging methods, especially UK with Matérn covariance functions, offer a promising alternative for genomic prediction, outperforming GBLUP under complex genetic architectures.
• The findings provide a theoretical foundation for understanding covariance structures in genomic prediction and suggest avenues for further methodological development.