Radial basis function regression methods for predicting quantitative traits using SNP markers.

Area of Science:

• Quantitative genetics
• Genomic prediction
• Machine learning in animal breeding

Background:

• Predicting genetic values for complex traits is hindered by cryptic gene interactions and non-additive effects.
• Linear models assuming additive effects may yield poor predictive ability when genotype-phenotype relationships are non-linear or non-additive.

Purpose of the Study:

• To evaluate the performance of non-parametric radial basis function (RBF) regression for predicting total genetic values.
• To compare RBF regression against a linear additive model (Bayes A) using simulations and real broiler data.

Main Methods:

• Simulations included varying heritability levels and gene action modes (additive, additive+dominance, epistasis).
• Two RBF models (SNP-specific weights - RBF I, common weights - RBF II) were tested.
• Performance was assessed using correlation and predictive mean-squared error (PMSE) on a testing set.

Main Results:

• RBF regression outperformed the Bayes A linear model in predicting genetic values when complex genotype-phenotype relationships were present.
• RBF I generally performed better than RBF II and identified relevant single nucleotide polymorphisms (SNPs) in a toy example.
• Extending linear models to include epistasis could improve accuracy.

Conclusions:

• Non-parametric RBF regression is a powerful tool for genomic prediction of complex traits, especially when non-linear and non-additive gene effects are involved.
• RBF models offer advantages over traditional linear models for capturing complex genetic architectures.
• RBF I demonstrates potential for identifying important SNPs in genomic prediction.