Multi Task deep learning for genomic predictions

Abstract

Genomic predictions have been recognized as a new promising technique in animal and plant breeding. Linear mixed model is a widely used statistical technique, but it may not be desirable for large training sets and number of molecular markers, because it is intensive in computation. Deep learning is a subfield of machine learning and it can be used for complex predictions on a large scale. Multi task deep learning (MT-DL) incorporates related tasks(labels or traits) into one learning process to enable the learning model to perform better than single task deep learning (ST-DL). I applied MT-DL to genotype by environment genomic predictions to predict the performances of breeding lines at multiple environments. I compared MT-DL with linear mixed model-based Bayesian genotype x environment method (BGGE) and separate genomic predictions on single environments with widely used rrBLUP, ridge regression and ST-DL using cross validations. Compared with rrBLUP, MT-DL and non-linear BGGE showed a moderate increase of 9.4 and 7.6%, respectively, ST-DL has a small increase of 5.4%, ridge regression had a similar prediction accuracy and linear BGGE had a small decrease of -2.0% for prediction accuracy. I also found that all methods including rrBLUP had an overfitting, this is likely because yield genomic predictions are complex and the data set used in this study are small. rrBLUP, ridge regression, ST-DL and MT-DL has similar overfitting. Difference between training and test set prediction accuracies was between 0.344 and 0. 387. Linear and nonlinear BGGE methods seem to have much worse overfitting than other methods. Difference between training and test set prediction accuracies were 0.429 and 0.472, respectively. I also discussed the potential applications of ST-DL and MT-DL in genomic predictions of hybrid crops such as maize