Application of machine learning methods on the genomic analysis for reproductive traits in Nelore cattle

• Grant number: 16/24227-2
• Support type: Scholarships in Brazil - Doctorate
• Effective date (Start): March 01, 2017
• Status: Discontinued
• Field of knowledge: Agronomical Sciences - Animal Husbandry
• Principal Investigator: Lucia Galvão de Albuquerque
• Grantee: Anderson Antonio Carvalho Alves
• Home Institution: Faculdade de Ciências Agrárias e Veterinárias (FCAV). Universidade Estadual Paulista (UNESP). Campus de Jaboticabal. Jaboticabal , SP, Brazil
• Associated research grant: 09/16118-5 - Genomic tools to genetic improvement of direct economic important traits in Nelore cattle, AP.TEM

Abstract

Evaluating and selecting animals based on genomic information has been an increasingly promising tendency in breeding programs. Nevertheless, principal approaches to perform genomic prediction involve parametric methodologies and in the mostly cases only assume additive variance for the markers effects, ignoring possible cryptic non-linear effects. Considering such effects may be important to improve the marker-based prediction ability for complex traits. Based on this, recently, there is a growing interest on semi and non-parametrical prediction methods. In this fashion, machine learning based methods such as Artificial Neural Networks (ANN), Random Forests (RF) and Support Vector Machines (SVM) could be an interesting alternative. Due its structure, ANN are able to learn non-linear relationships between molecular markers and phenotype in a adaptive way, without to specify a model. However, due its computational cost, such methodology has been few explored in commercial herds data. A way to overcome the computing and statistical constraints is through bayesian regularization. On the other hand, RF combines bootstrapping and decision trees techniques widely applied for classification and regression tasks. Another flexible method for these proposes known as SVM uses kernel functions for fitting the data. The aim of this study will be to assess the impact from the use of different machine learning methods on genomic prediction accuracy for reproductive traits in Nelore Cattle. The following traits will be included in this study: First Calving Age (FCA), Early Pregnancy Occurrence (EPO), Days to Calving (DC), Scrotal Perimeter (EP). The phenotypes will be corrected for the fixed effects considering animals from same herd, birth season and management group. Different neural networks architectures will be evaluated, as well as the input information structure in order to assess the prediction ability for the studied traits. The results will be compared by cross-validation to those obtained by benchmark marker-based genomic prediction models as Bayesian LASSO and GBLUP. Comparison criteria will be the Pearson correlation between observed and predicted phenotypes, besides the variance prediction error in the validation groups. Additionally, RF performance will be used to identify candidate regions that could be associated with the studied traits.