B0414
Title: Genome and phenome statistical models and methods for prediction
Authors: Osval Montesinos-Lopez - Universidad de Colima (Mexico)
Paulino Perez-Rodriguez - Colegio de Postgraduados (Mexico)
Reka Howard - University of Nebraska - Lincoln (United States)
Jaime Cuevas - Universidad de Quintana Roo (Mexico)
Diego Jarquin - University of Nebraska - Lincoln (United States)
Abelardo Montesins-Lopez - Universidad de Guadalajara (Mauritius)
Jose Crossa - University of Nebraska-Lincoln (United States) [presenting]
Abstract: In the last years genome and phenome models have been developed for the prediction of unobserved individuals using dense molecular markers and high throughput phenotype (HTP) informations. Statistical models include single and multi-traits, and single and multi-environments as well as several HTP and near infrared spectroscopy (NIR) information with the objective of increase the prediction accuracy of grain yield, and other traits on unobserved individuals. Increase in prediction accuracy over the genomic best linear unbiased predictor (GBLUP) was achieved by means of the Gaussian kernel model including genomic by environments interaction (GE). A Bayesian multi-trait multi-environment model can efficiently exploit correlated traits and environments and thus increasing the prediction accuracy of about 10\% over the single trait, single environment model. Bayesian models requires intense computational resources. Deep Machine Learner (DL) models with densely connected network architecture were developed with the objective of using less computing resources and accommodate extensive data sets. Under certain circumstances DL were competitive with other well established models. Although implementing the multi-trait DL models is feasible and practical in the genomic prediction context it is challenging due to the large number of hyper-parameters involved. Deep Kernel method has been studied and results compared with other kernel methods.
