Predicting complex quantitative traits with Bayesian neural networks: a case study with Jersey cows and wheat.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Complex phenotypes often involve non-linear genetic associations not captured by traditional parametric models.
• Existing methods like Bayesian linear regression assume additive inheritance, ignoring gene interactions and non-linear effects.
• Artificial neural networks (ANNs) offer a powerful alternative, acting as universal approximators for complex functions.

Purpose of the Study:

• To investigate the utility of various Bayesian artificial neural network (ANN) architectures for predicting phenotypes using genomic data.
• To compare the predictive performance of ANNs against linear models in complex trait prediction.
• To assess the effectiveness of ANNs in handling high-dimensional genomic information and non-linear relationships.

Main Methods:

• Bayesian ANN architectures were applied to two datasets: milk production in Jersey cows and yield in wheat lines.
• Predictor variables included pedigree and single nucleotide polymorphism (SNP) marker data.
• ANN models were trained and their predictive abilities compared to a linear model.

Main Results:

• Predictive ability for milk production traits in Jersey cows was low with pedigree data but improved with SNP data.
• Wheat yield prediction showed higher accuracy than cow milk production, attributed to predicting a mean trait.
• Non-linear ANNs consistently outperformed linear models in predictive ability for both datasets, particularly in wheat.

Conclusions:

• Artificial neural networks (ANNs) show promise for predicting complex traits using high-dimensional genomic data.
• ANNs can adaptively capture non-linear genetic effects, which is crucial for accurate phenotype prediction.
• The ability of ANNs to handle complex, non-linear relationships makes them valuable tools in genomic prediction.