Prediction of Indigenous Chicken Genomic Breeding Values and Performance using Machine Learning

Abstract

Genomic selection (GS) is a new breeding strategy that is rapidly becoming the method of choice of selection. It predicts the phenotypes of quantitative traits based on genome-wide markers of genotypes using conventional predictive models. However. these conventional predictive models are faced with a statistical challenge related to high dimensionality of marker data and typically make strong assumptions related to linear regression analysis. They also do not capture the complex, non-linear relationships Within genotypes and between genotypes and phenotypes. Machine learning has been reported to be a promising tool, which accounts for these shortcomings. This study aimed at detemiining the adoption of machine leaming in genomic prediction and performance estimation of indigenous chicken (IC). To compare predictive ability of genomic models, DeepGS, ridge-regression best linear unbiased prediction (RR-BLUP), Ensemble and artificial neural network (ANN) models were adopted in predicting body weight (BW) of IC based on genome-wide markers. The Pearson correlation coefficient (PCC) results for the four models were 0.891, 0.889, 0.892 and 0.812 respectively. All models were not significant difference (p>0.05) from each other, therefore, all the four models can be used in complement. For perfonnance prediction, mean normalized discounted cumulative gain value (MNDCGV) were adopted. The MNDCGV results showed that the accuracy of genomic estimated breeding values (GEBVs) estimated using DeepGS RR-BLUP and Ensemble were 0.75~0.78, 0.66~O.76 and O.76~0.79 respectively. Thus, the Ensemble and DeepGS model outperformed the RR-BLUP model by a significant margin (P<0.05), therefore they can be used as a supplement to RR-BLUP. To evaluate the extent of non-linearity among explanatory variables within genotypes and between genotypes and phenotypes, a multilayer perceptron ANN was adopted. Mean absolute error (MAE) and PCC were used to measure the ANN perfonnance. The results showed that the neural network with one hidden layer containing 10 neurons yielded high PCC value of 0.86 and MSE of 2.98E-3. Further increase of network dimension to 16 and 32 neurons the performance decreased to 0.67 and 0.27 for PCC and MAE increased to 7.73E-2 and 7.60E-2 respectively. Thus, model with 10 neurons is enough to handle non-linearity of data set thus chosen as the best non- linear model. This study concludes that DeepGS, Ensemble, RR-BLUP and ANN models can be used interchangeably in making phenotypic predictions, also the Ensemble and DeepGS model are appropriate in phenotypic ranking of individuals. vi