Prediction Of Pig Breeding Phenotype And Screening Of Gene Chip Loci Based On Machine Learning

Pig breeding is an important branch in the agricultural field,and its development plays a vital role in improving the production efficiency,quality,and supply of pork.Genomic prediction based on solving linear mixed model equations was the most common method for predicting breeding values or phenotypic performance of economic traits in livestock.However,to further improve the performance of genome prediction,nonlinear methods were considered as an alternative with great potential.Machine learning methods were rapidly developing and had shown excellent phenotype prediction capabilities in livestock breeding.This paper focused on the aspects of linear and nonlinear methods for genome prediction as follows:(1)A total of eight important production traits in pig breeding were compared for prediction accuracy between the linear model and the nonlinear model using the PIC public dataset from the world’s leading pig breeding company and the national pig core group dataset in Chifeng,North China.Different nonlinear machine learning models were constructed for pig phenotype prediction based on pig gene chip data to study the feasibility and reliability of genome prediction using nonlinear models.These included random forest,support vector machine,extreme gradient boosting trees,and convolutional neural networks for deep learning.The results showed that for traits T1,T2,T3,and T5 in the PIC dataset and the average daily gain in the Chifeng dataset,support vector machine model had higher prediction accuracy than the linear mixed method.However,for trait T4 in the PIC dataset,the total number of piglets born in the Chifeng dataset,and the psoas muscle depth,the accuracy of the support vector machine model was slightly lower than that of the linear mixed method.Nonlinear machine learning models showed predictive accuracy exceeding linear mixed methods on 60%of the datasets.This showed that support vector machine models can be very helpful for pig breeding.(2)Four feature selection methods,including recursive feature elimination,extreme gradient boosting machine,random forest,and genome-wide association study,were used to screen and reduce the dimensionality of high-dimensional pig genome markers to reduce the cost of pig gene chips.Based on the support vector machine model with the best prediction effect,a small number of features selected by several feature selection methods were used for prediction,and the influence of the selected features on the prediction effect and computing performance was explored.The results showed that the extreme gradient boosting machine model reduced the number of genomic markers to one-twentieth,and the prediction accuracy of a small number of features selected by it was the best among the four models.Among the common feature selection methods,the extreme gradient boosting machine method was found to be more suitable for feature screening of genetic data.(3)Integrated Extreme Gradient Boosting Machine and Support Vector Machine models were developed into a new tool,which could realize genomic site screening and phenotype prediction functions as SNPkey,making genome prediction phenotype work more time-saving and convenient.