Predicting expected progeny difference for marbling score in Angus cattle using artificial neural networks and...

Area of Science:

• Animal Genomics
• Bioinformatics
• Machine Learning in Agriculture

Background:

• Artificial neural networks (ANN) mimic the human brain for parallel data processing and can model complex, nonlinear relationships.
• ANN with Bayesian regularization have previously shown superior performance over linear models in predicting crop and dairy yields.
• The potential of ANN to improve the accuracy of molecular breeding values for genomic selection warrants investigation.

Purpose of the Study:

• To evaluate the accuracy of artificial neural networks (ANN) for predicting the expected progeny difference (EPD) of marbling score in Angus cattle.
• To compare the performance of various ANN architectures against established linear Bayesian regression models.

Main Methods:

• Exploration of diverse ANN architectures, including variations in training algorithms, activation functions, and hidden layer neuron counts.
• Implementation of ANN with Bayesian regularization (BRANN) for prediction.
• Comparison with BayesCpC models, which involve marker selection and effect estimation under an additive inheritance assumption.

Main Results:

• ANN with Bayesian regularization demonstrated comparable prediction accuracy and sum of squared errors to the BayesCpC method.
• For a 3K-SNP panel, BRANN achieved prediction accuracies ranging from 0.776 to 0.807, while BayesCpC yielded 0.776.
• Using a selected 700-SNP panel, BRANN's accuracy ranged from 0.842 to 0.858, compared to 0.863 for BayesCpC. ANN using scaled conjugate gradient back-propagation showed lower accuracy.

Conclusions:

• Artificial neural networks with Bayesian regularization are effective for predicting additive genetic values, performing on par with linear Bayesian models.
• ANN serve as valuable universal approximators for complex functions relevant to animal breeding and genomic selection.