Deep Learning Algorithms for Complex Traits Genomic Prediction

Abstract

The underlying perception of genomic selection (GS) is to use genome-wide from DNA sequence (“SNP markers”) along with phenotypes from an observed population to make prediction for the phenotypic outcomes of untested individuals in crop and livestock breeding programs. GS was firstly described by Meuwissen et al.(2001) in dairy cattle to identify genetically superior animals at an early age. The aim was to capture specific genes across the whole genome that are associated with desired traits. The major challenge in using GS programs is to predict the effect of many SNP markers using phenotypic information from a few individuals (aka small n big p problem, or p >> n). Many approaches including naïve and scaled elastic net, ridge regression BLUP Bayesian approaches (BayesA, BayesB, BayesCπ, BayesDπ) LASSO, Support Vector Regression have been conducted to address the small n big p (aka, p >> n) problem. These methods all perform well for (p>>n) by using linear approximation to set a functional relationship between genotypes and phenotypes. However, these methods may not fully capture non-linear effects which are possible to be crucial for complex traits. To deal with this limitation, many methods including neural networks (NN) were recommended to cover non-linearity for GS. Artificial NNs (ANNs) for GS was first presented by Okut et al. (2011) who establish a fully connected regularized multi-layer ANN (MLANN) comprising one hidden layer to predict the body mass index (BMI) in mice using dense molecular markers. Since then, rather complex ANNs approaches have been applied including deep learning (DL) networks. The different DL algorithms have their own advantages to deal with specific problems in complex trait GS. Four different major classes of DL approaches such as fully connected deep learning artificial neural networks (DL-MLANN), recurrent neural networks (RNN), convolutional neural networks (CNN) and long-short term memory (LSTM) and some variation of these network architectures will be summarized here.