Predictions of Indigenous Chicken Phenotypes from Genotypes: Comparison Between Machine Learning and conventional Linear Models

Genomic selection is a new breeding strategy which is rapidly becoming the method of choice of selection. It is useful in predicting the phenotypes of quantitative traits based on genome-wide markers of genotypes using conventional predictive models such as ridge regression BLUP. However, these conventional predictive models are faced with a statistical challenge related to the high dimensionality of marker data, inter and intra-allelic interactions and typically make strong assumptions. Machine learning models can be used as an alternative in the prediction of phenotypes due to their ability to address this challenges. Therefore, the aim of this study was to compare the predictive ability of machine learning using deep convolutional neural network (DeepGS), conventional neural network (Artificial neural network), conventional statistical predictive model ridge regression best linear unbiased (RR-BLUP) model and combination of DeepGS and RR-BLUP(Ensemble model) in predicting body weight (BW) of indigenous chicken based on genome-wide markers. The pearson correlation coefficient (PCC) results from this study for the four models were 0.891,0.889, 0.892 and 0.812 for DeepGS, RR-BLUP, Ensemble and ANN. This showed that DeepGS did not yield significant difference (p>0.05) from the other models, therefore, it can be used in complement to the commonly used conventional models. For individuals with higher phenotypic values, the PCC results showed a drastic decrease in the performance of DeepGS, rrBLUP, Ensemble and ANN from 0.891, 0.889, 0.892, 0.845 to 0.315, 0.466, 0.342, 0.518 respectively. Therefore, more effort should be put on individuals with higher phenotypic values.