Deep kernel of genomic and near infrared predictions in multi-environment breeding trials

In genomic prediction deep learning artificial neural network are part of machine learning methods that incorporate parametric, non-parametric and semi-parametric statistical models. Kernel methods are seeing more flexible, and easier to interpret than neural networks. Kernel methods used in genomic predictions comprise the linear genomic best linear unbiased predictor (GBLUP) kernel (GB) and the Gaussian kernel (GK). These kernels have being used with two statistical models, single environment and genomic × environment (GE) models. Recently near infrared spectroscopy (NIR) has being used as phenotype method for prediction of unobserved line performance in plant breeding trials. In this study, we used a non-linear Arc-cosine kernel (AK) that emulates deep learning artificial neural network. We compared AK prediction accuracy with GB and GK kernel methods in four genomic data sets one of them including also pedigree (ABLUP) and NIR (NBLUP) information. Results show that for all four data sets AK and GK kernels gave higher prediction accuracy than the linear GB kernel for single environment as well as GE multi-environment models. In addition, AK gave similar or slightly higher prediction accuracy than the GK kernel.