Genomic Selection with Deep Neural Networks

Abstract

<p>Reduced costs for DNA marker technology has generated a huge amount of molecular</p> <p>data and made it economically feasible to generate dense genome-wide marker maps of lines</p> <p>in a breeding program. Increased data density and volume has driven an exploration of</p> <p>tools and techniques to analyze these data for cultivar improvement. Data science theory</p> <p>and application has experienced a resurgence of research into techniques to detect or ”learn”</p> <p>patterns in noisy data in a variety of technical applications. Several variants of machine</p> <p>learning have been proposed for analyzing large DNA marker data sets to aid in pheno-</p> <p>type prediction and genomic selection. Here, we present a review of the genomic prediction</p> <p>and machine learning literature. We apply deep learning techniques from machine learn-</p> <p>ing research to six phenotypic prediction tasks using published reference datasets. Because</p> <p>regularization frequently improves neural network prediction accuracy, we included regular-</p> <p>ization methods in the neural network models. The neural network models are compared to</p> <p>a selection of regularized Bayesian and linear regression techniques commonly employed for</p> <p>phenotypic prediction and genomic selection. On three of the phenotype prediction tasks,</p> <p>regularized neural networks were the most accurate of the models evaluated. Surprisingly,</p> <p>for these data sets the depth of the network architecture did not affect the accuracy of the</p> <p>trained model. We also find that concerns about the computer processing time needed to</p> <p>train neural network models to perform well in genomic prediction tasks may not apply when</p> <p>Graphics Processing Units are used for model training.</p>