Background: Several conventional genomic Bayesian (or no Bayesian) prediction methods have been proposed including the standard additive genetic effect model for which the variance components are estimated with mixed model equations. In recent years, deep learning (DL) methods have been considered in the context of genomic prediction. The DL methods are nonparametric models providing flexibility to adapt to complicated associations between data and output with the ability to adapt to very complex patterns.Main body: We review the applications of deep learning (DL) methods in genomic selection (GS) to obtain a meta-picture of GS performance and highlight how these tools can help solve challenging plant breeding problems. We also provide general guidance for the effective use of DL methods including the fundamentals of DL and the requirements for its appropriate use. We discuss the pros and cons of this technique compared to traditional genomic prediction approaches as well as the current trends in DL applications.Conclusions: The main requirement for using DL is the quality and sufficiently large training data. Although, based on current literature GS in plant and animal breeding we did not find clear superiority of DL in terms of prediction power compared to conventional genome based prediction models. Nevertheless, there are clear evidences that DL algorithms capture nonlinear patterns more efficiently than conventional genome based. Deep learning algorithms are able to integrate data from different sources as is usually needed in GS assisted breeding and it shows the ability for improving prediction accuracy for large plant breeding data. It is important to apply DL to large training-testing data sets.

Genomic selection is revolutionizing plant breeding. However, still lacking are better statistical models for ordinal phenotypes to improve the accuracy of the selection of candidate genotypes. For this reason, in this paper we explore the genomic based prediction performance of two popular machine learning methods: the Multi Layer Perceptron (MLP) and support vector machine (SVM) methods vs. the Bayesian threshold genomic best linear unbiased prediction (TGBLUP) model. We used the percentage of cases correctly classified (PCCC) as a metric to measure the prediction performance, and seven real data sets to evaluate the prediction accuracy, and found that the best predictions (in four out of the seven data sets) in terms of PCCC occurred under the TGLBUP model, while the worst occurred under the SVM method. Also, in general we found no statistical differences between using 1, 2 and 3 layers under the MLP models, which means that many times the conventional neuronal network model with only one layer is enough. However, although even that the TGBLUP model was better, we found that the predictions of MLP and SVM were very competitive with the advantage that the SVM was the most efficient in terms of the computational time required.

Multiple-trait experiments with mixed phenotypes (binary, ordinal and continuous) are not rare in animal and plant breeding programs. However, there is a lack of statistical models that can exploit the correlation between traits with mixed phenotypes in order to improve prediction accuracy in the context of genomic selection (GS). For this reason, when breeders have mixed phenotypes, they usually analyze them using univariate models, and thus are not able to exploit the correlation between traits, which many times helps improve prediction accuracy. In this paper we propose applying deep learning for analyzing multiple traits with mixed phenotype data in terms of prediction accuracy. The prediction performance of multiple-trait deep learning with mixed phenotypes (MTDLMP) models was compared to the performance of univariate deep learning (UDL) models. Both models were evaluated using predictors with and without the genotype × environment (G×E) interaction term (I and WI, respectively). The metric used for evaluating prediction accuracy was Pearson’s correlation for continuous traits and the percentage of cases correctly classified (PCCC) for binary and ordinal traits. We found that a modest gain in prediction accuracy was obtained only in the continuous trait under the MTDLMP model compared to the UDL model, whereas for the other traits (1 binary and 2 ordinal) we did not find any difference between the two models. In both models we observed that the prediction performance was better for WI than for I. The MTDLMP model is a good alternative for performing simultaneous predictions of mixed phenotypes (binary, ordinal and continuous) in the context of GS.

Modern agriculture uses hyperspectral cameras that provide hundreds of reflectance data at discrete narrow bands in many environments. These bands often cover the whole visible light spectrum and part of the infrared and ultraviolet light spectra. With the bands, vegetation indices are constructed for predicting agronomically important traits such as grain yield and biomass. However, since vegetation indices only use some wavelengths (referred to as bands), we propose using all bands simultaneously as predictor variables for the primary trait grain yield; results of several multi-environment maize (Aguate et al. in Crop Sci 57(5):1–8, 2017) and wheat (Montesinos-López et al. in Plant Methods 13(4):1–23, 2017) breeding trials indicated that using all bands produced better prediction accuracy than vegetation indices. However, until now, these prediction models have not accounted for the effects of genotype × environment (G × E) and band × environment (B × E) interactions incorporating genomic or pedigree information. In this study, we propose Bayesian functional regression models that take into account all available bands, genomic or pedigree information, the main effects of lines and environments, as well as G × E and B × E interaction effects. The data set used is comprised of 976 wheat lines evaluated for grain yield in three environments (Drought, Irrigated and Reduced Irrigation). The reflectance data were measured in 250 discrete narrow bands ranging from 392 to 851 nm (nm). The proposed Bayesian functional regression models were implemented using two types of basis: B-splines and Fourier. Results of the proposed Bayesian functional regression models, including all the wavelengths for predicting grain yield, were compared with results from conventional models with and without bands. We observed that the models with B × E interaction terms were the most accurate models, whereas the functional regression models (with B-splines and Fourier basis) and the conventional models performed similarly in terms of prediction accuracy. However, the functional regression models are more parsimonious and computationally more efficient because the number of beta coefficients to be estimated is 21 (number of basis), rather than estimating the 250 regression coefficients for all bands. In this study adding pedigree or genomic information did not increase prediction accuracy.

Deep learning (DL) is revolutionizing the development of artificial intelligence systems. For example, before 2015, humans were better than artificial machines at classifying images and solving many problems of computer vision (related to object localization and detection using images), but nowadays, artificial machines have surpassed the ability of humans in this specific task. This is just one example of how the application of these models has surpassed human abilities and the performance of other machine‐learning algorithms. For this reason, DL models have been adopted for genomic selection (GS). In this article we provide insight about the power of DL in solving complex prediction tasks and how combining GS and DL models can accelerate the revolution provoked by GS methodology in plant breeding. Furthermore, we will mention some trends of DL methods, emphasizing some areas of opportunity to really exploit the DL methodology in GS; however, we are aware that considerable research is required to be able not only to use the existing DL in conjunction with GS, but to adapt and develop DL methods that take the peculiarities of breeding inputs and GS into consideration.

Genomic selection (GS) is transforming plant and animal breeding, but its practical implementation for complex traits and multi-environmental trials remains challenging. To address this issue, this study investigates the integration of environmental information with genotypic information in GS. The study proposes the use of two feature selection methods (Pearson’s correlation and Boruta) for the integration of environmental information. Results indicate that the simple incorporation of environmental covariates may increase or decrease prediction accuracy depending on the case. However, optimal incorporation of environmental covariates using feature selection significantly improves prediction accuracy in four out of six datasets between 14.25% and 218.71% under a leave one environment out cross validation scenario in terms of Normalized Root Mean Squared Error, but not relevant gain was observed in terms of Pearson´s correlation. In two datasets where environmental covariates are unrelated to the response variable, feature selection is unable to enhance prediction accuracy. Therefore, the study provides empirical evidence supporting the use of feature selection to improve the prediction power of GS.