Subset selection of markers for the genome-enabled prediction of genetic values using radial basis function neural networks

Abstract

This paper aimed to evaluate the effectiveness of subset selection of markers for genome-enabled prediction of genetic values using radial basis function neural networks (RBFNN). To this end, an F1 population derived from the hybridization of divergent parents with 500 individuals genotyped with 1000 SNP-type markers was simulated. Phenotypic traits were determined by adopting three different gene action models – additive, additive-dominant, and epistatic, representing two dominance situations: partial and complete with quantitative traits having a heritability (h2) of 30 and 60%; traits were controlled by 50 loci, considering two alleles per locus. Twelve different scenarios were represented in the simulation. The stepwise regression was used before the prediction methods. The reliability and the root mean square error were used for estimation using a fivefold cross-validation scheme. Overall, dimensionality reduction improved the reliability values for all scenarios, specifically with h2 =30 the reliability value from 0.03 to 0.59 using RBFNN and from 0.10 to 0.57 with RR-BLUP in the scenario with additive effects. In the additive dominant scenario, the reliability values changed from 0.12 to 0.59 using RBFNN and from 0.12 to 0.58 with RR-BLUP, and in the epistasis scenarios, the reliability values changed from 0.07 to 0.50 using RBFNN and from 0.06 to 0.47 with RR-BLUP. The results showed that the use of stepwise regression before the use of these techniques led to an improvement in the accuracy of prediction of the genetic value and, mainly, to a large reduction of the root mean square error in addition to facilitating processing and analysis time due to a reduction in dimensionality.

Genome‐enabled prediction through machine learning methods considering different levels of trait complexity

“…The adverse effect of lower heritability on selective accuracy in several scenarios has been proven in previous studies that used the RR‐BLUP (Coutinho et al., 2018; Moura et al., 2019), BayesB methods (Moura et al., 2019), BO (Ghafouri‐Kesbi et al., 2017), RF (Ghafouri‐Kesbi et al., 2017), RBF (Sant'Anna et al., 2020), and MLP methods (Coutinho et al., 2018). Guo et al.…”

“…Different methods capable of addressing problems related to data dimensionality have been used, such as ridge regression‐best linear unbiased predictor (RR‐BLUP; Endelman, 2011) and Bayesian methods (Meuwissen et al., 2001). More recent approaches based on machine learning have also been adopted through the methods of radial basis function (RBF), multilayer perceptron (MLP; Ehret et al., 2015; Sant'Anna et al., 2020; Silva et al., 2014), and decision trees (DT) with their refinements boosting (BO), random forest (RF), and bagging (BA; de Sousa et al., 2021).…”

“…Machine learning‐based methods also have other properties that make them potentially more attractive to GWS (Ferreira et al., 2018; Sant'Anna et al., 2020; Silva et al., 2014; de Sousa et al., 2021). This is because all markers have a chance to contribute to the adjustment of the model, including those with weak effects and highly correlated and interacting markers.…”

“…This is because all markers have a chance to contribute to the adjustment of the model, including those with weak effects and highly correlated and interacting markers. Additionally, these methods allow complex interactions between markers to be easily included, such as in nonadditive models with dominant and epistatic effects, and do not make distributive assumptions about the predictor variables (I. C. de Ferreira et al., 2018; Sant'Anna et al., 2020; Silva et al., 2014; de Sousa et al., 2021).…”

Artificial Neural Networks in the Prediction of Genetic Merit to Flowering Traits in Bean Cultivars

“…In addition, this approach does not require any assumptions about the distribution of phenotypic values as the statistical methods do. ANNs have been used successfully in several breeding studies to predict the genetic merit using simulated [15,16] and real data [17][18][19]. Overall, these studies show that the application of ANN in GS presents great potential for capturing complex interactions since the accuracy values and the bias are, respectively, higher and lower compared with those obtained through traditional GS methodologies (for example, G-BLUP).…”