Genome-enabled prediction of genetic values using radial basis function neural networks.

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Genomic selection (GS) is crucial in modern animal and plant breeding.
• High-density molecular markers enable advanced prediction models.
• Various regression models are employed for quantitative trait prediction in GS.

Purpose of the Study:

• To evaluate the predictive performance of radial basis function neural networks (RBFNN) and reproducing kernel Hilbert spaces (RKHS) regression compared to Bayesian LASSO.
• To assess the utility of non-linear models for genomic prediction using dense molecular markers.
• To investigate the impact of epistatic effects and redundant predictors on prediction accuracy.

Main Methods:

• Comparison of three regression models: Bayesian LASSO, RKHS regression, and RBFNN.
• Application to simulated data and real maize lines genotyped with 55,000 markers.
• Evaluation across multiple trait-environment combinations.

Main Results:

• All three models demonstrated comparable overall prediction accuracy.
• RKHS and RBFNN showed a slight, consistent advantage over the additive Bayesian LASSO model.
• RKHS and RBFNN models captured epistatic effects in simulated data.

Conclusions:

• Non-linear models like RKHS and RBFNN offer slight improvements in genomic prediction accuracy over linear models.
• While capable of capturing epistasis, non-linear models can be sensitive to the inclusion of redundant predictors, potentially reducing accuracy.