A Sparse Convolutional Predictor with Denoising Autoencoders for Phenotype Prediction

Abstract

Phenotype prediction has been widely conducted in many areas to help understand disease risks and susceptibility, and improve the breeding cycles of plants and animals. Most methods of phenotype prediction are based on regularized statistical approaches which only consider linear relationships among genetic features. Deep learning based methods have been recently reported to nicely address regression problems in high dimensional data in genomic studies. To explore deep learning for phenotype prediction, we propose a deep learning regression model, called Sparse Convolutional Predictor with Denoising Autoencoders (SCP_DAE), to predict quantitative traits. We constructed SCP_DAE by utilizing a convolutional layer that can extract correlation or linkage patterns in the genotype data and applying a sparse weight matrix resulted from the L1 regularization to handle high dimensional genotype data. To learn efficient and compressed hidden representations of genotype data, we pre-trained the convolutional layer and the first fully connected layer in SCP_DAE using denoising autoencoders. These pre-trained layers were then fine-tuned to improve its performance of the SCP_DAE model for phenotype prediction. We comprehensively evaluated our proposed method on a yeast dataset which contains well assayed genotype profiles and quantitative traits. Our results showed that the proposed SCP\_DAE method significantly outperforms regularized statistical approaches and similar deep learning models without pre-trained weights.