Evaluation of the efficiency of artificial neural networks for genetic value prediction

Abstract

ABSTRACT. Artificial neural networks have shown great potential when applied to breeding programs. In this study, we propose the use of artificial neural networks as a viable alternative to conventional prediction methods. We conduct a thorough evaluation of the efficiency of these networks with respect to the prediction of breeding values. Therefore, we considered eight simulated scenarios, and for the purpose of genetic value prediction, seven statistical parameters in addition to the phenotypic mean in a network designed as a multilayer perceptron. After an evaluation of different network configurations, the results demonstrated the superiority of neural networks compared to estimation procedures based on linear models, and indicated high predictive accuracy and network efficiency.

Artificial neural networks compared with Bayesian generalized linear regression for leaf rust resistance prediction in Arabica coffee

“…It is believed that the ANN approach, due to its different neurons in hidden layers, can capture linear and nonlinear relationships between the response variable and the values assigned to molecular markers (Braga et al, 2011). This means that they were capable of detecting the effects of dominance and epistasis that are neglected by other methodologies (Heslot et al, 2012;Silva et al, 2014Silva et al, , 2016.…”

“…Percentage agreements were of 89 and 100% for adaptability, and of 78 and 100% for stability, considering favorable and unfavorable environments, respectively. Silva et al (2014Silva et al ( , 2016 concluded that ANNs are efficient in predicting values and genetic gain in simulated trials under randomized block design. Regarding the classification studies, Sant'Anna et al (2015) showed that neural networks had results superior to those obtained by discriminant analysis in the classification of simulated populations.…”

“…In the genetic context, it should be taken into account that BGLR approach uses a linear model in which the predicted value -being function only of favorable allele dose -predicts the additive genetic value with higher mean squared error, or lower genotypic reliability, expressed by the r 2 , of the model because of the effects assigned to dominance and epistatic interactions disregarded in this model (Cruz, 2012). However, when a computational intelligence approach as ANN is used, such information and other important information that the breeder has about the genotype in study constitute a relevant input to be considered during the training processes for increasing the efficiency of the selection process, and it becomes possible to capture different relationships between markers and phenotypes (Bureau et al, 2005;Heslot et al, 2012;Silva et al, 2014Silva et al, , 2016.…”

“…For the network training (obtainment of weights), the set of observations was divided in two parts: training and validation, as commonly adopted in the literature (Silva et al, 2014(Silva et al, , 2016Sant'Anna et al, 2015). The first, denoted by training set, consisted of 160 individuals taken at random.…”

“…Such approach has been successfully applied to solve several problems related to genetics (Nascimento et al, 2013;Silva et al, 2014Silva et al, , 2016Sant'Anna et al, 2015).…”