Gene Expression Prediction Using Deep Neural Networks

Abstract

In the field of molecular biology, gene expression is a term that encompasses all the information contained in an organism’s genome. Although, researchers have developed several clinical techniques to quantitatively measure the expressions of genes of an organism, they are too costly to be extensively used. The NIH LINCS program revealed that human gene expressions are highly correlated. Further research at the University of California, Irvine (UCI) led to the development of D-GEX, a Multi Layer Perceptron (MLP) model that was trained to predict unknown target expressions from previously identified landmark expressions. But, bowing to hardware limitations, they had split the target genes into different sets and constructed separate models to profile the whole genome. This paper proposes an alternative solution using a combination of deep autoencoder and MLP to overcome this bottleneck and improve the prediction performance. The microarray based Gene Expression Omnibus (GEO) dataset was employed to train the neural networks. Experimental result shows that this new model, abbreviated as E-GEX, outperforms D-GEX by 16.64% in terms of overall prediction accuracy on GEO dataset. The models were further tested on an RNA-Seq based 1000G dataset and E-GEX was found to be 49.23% more accurate than D-GEX.