Genome-based prediction of Bayesian linear and non-linear regression models for ordinal data

Area of Science:

• Quantitative Genetics
• Machine Learning in Breeding
• Statistical Genomics

Background:

• Genomic selection (GS) utilizes genomic information for predicting breeding values.
• Existing models often focus on continuous traits, with less development for ordinal or categorical responses.
• Artificial neural networks offer potential for enhanced prediction accuracy in complex trait prediction.

Purpose of the Study:

• To propose a novel Bayesian Regularized Neural Network (BRNNO) model tailored for ordinal data.
• To evaluate the performance of the BRNNO model in genomic prediction applications.
• To compare the BRNNO model against established methods like the Bayesian Ordered Probit Model (BOPM).

Main Methods:

• Development and implementation of a Bayesian Regularized Neural Network (BRNNO) model.
• Utilized a Bayesian framework with data augmentation for computational efficiency.
• Employed Gibbs Maximum a Posteriori and Generalized EM algorithms, coded in C and R.

Main Results:

• The BRNNO model demonstrated superior genomic-based prediction performance compared to the BOPM.
• The model was successfully tested on two real-world maize datasets for disease resistance traits (Septoria and GLS).
• The Bayesian approach facilitated robust parameter estimation and prediction for ordinal trait data.

Conclusions:

• The proposed BRNNO model is effective for modeling ordinal responses in genomic selection.
• BRNNO offers improved prediction accuracy for complex traits in animal and plant breeding.
• This study highlights the utility of advanced machine learning techniques within a Bayesian framework for genetic prediction.