Bayesian neural networks with variable selection for prediction of genotypic values.

Area of Science:

• Genetics
• Machine Learning
• Animal Breeding

Background:

• Estimating genetic components of complex phenotypes is challenging due to numerous allele effects and limited data.
• Traditional linear methods predict additive genetic effects but struggle with non-additive genetic effects.
• Machine learning, particularly neural networks, can model complex non-additive relationships.

Purpose of the Study:

• To develop and evaluate NetSparse, a novel method using Bayesian neural networks with variable selection for predicting individual genotypic values.
• To incorporate non-additive genetic effects into the prediction of complex traits.
• To assess NetSparse's accuracy compared to existing genomic prediction methods.

Main Methods:

• Developed NetSparse, a method combining Bayesian neural networks with variable selection.
• Simulated populations with diverse phenotypic models to test prediction accuracy.
• Compared NetSparse against genomic best linear unbiased prediction (GBLUP), BayesB, and their dominance variants.

Main Results:

• NetSparse demonstrated higher accuracy (2-28 percentage points) than reference methods for small numbers of quantitative trait loci (QTL).
• For dominance and epistatic effects, NetSparse showed improved accuracy (0.0-3.9 percentage points), except in extreme overdominance scenarios.
• Reference methods explicitly modeling dominance outperformed NetSparse by 6 percentage points in extreme overdominance cases.

Conclusions:

• Bayesian neural networks with variable selection show promise for predicting the genetic component of complex traits in animal breeding.
• NetSparse's performance is robust across various genetic models.
• High computational costs associated with NetSparse may limit its practical application.