Deep kernel and deep learning for genomic-based prediction

Deep learning (DL) is a promising method in the context of genomic prediction for selecting individuals early in time without measuring their phenotypes. iI this paper we compare the performance in terms of genome-based prediction of the DL method, deep kernel (arc-cosine kernel, AK) method, Gaussian kernel (GK) method and the conventional kernel method (Genomic Best Linear Unbiased Predictor, GBLUP, GB). We used two real wheat data sets for the benchmarking of these methods. We found that the GK and deep kernel AK methods outperformed the DL and the conventional GB methods, although the gain in terms of prediction performance of AK and GK was not very large but they have the advantage that no tuning parameters are required. Furthermore, although AK and GK had similar genomic-based performance, deep kernel AK is easier to implement than the GK. For this reason, our results suggest that AK is an alternative to DL models with the advantage that no tuning process is required.

Bibliographic Details

Main Authors:	Crossa, Jose, Martini, Johannes, Gianola, Daniel, Pérez-Rodríguez, Paulino, Burgueño, Juan, Singh, Ravi, Juliana, Philomin, Montesinos-López, Osval A., Cuevas, Jaime
Other Authors:	Shrestha, Rosemary
Format:	Experimental data, Genotypic data, Phenotypic data biblioteca
Language:	English
Published:	CIMMYT Research Data & Software Repository Network 2019
Subjects:	Agricultural Sciences, Agricultural research, Wheat, Triticum aestivum, Deep learning, Deep kernel, Genomic best linear unbiased predictor,
Online Access:	https://hdl.handle.net/11529/10548273
Tags:	Add Tag No Tags, Be the first to tag this record!