Encoding Gene Expression Using Deep Autoencoders for Expression Inference

Abstract

Gene expression of an organism contains all the information that characterises its observable traits. Researchers have invested abundant time and money to quantitatively measure the expressions in laboratories. On account of such techniques being too expensive to be widely used, the correlation between expressions of certain genes was exploited to develop statistical solutions. Pioneered by the National Institutes of Health Library of Integrated Network-Based Cellular Signature (NIH LINCS) program, expression inference techniques has many improvements over the years. The Deep Learning for Gene expression (D-GEX) project by University of California, Irvine approached the problem from a machine learning perspective, leading to the development of a multi-layer feedforward neural network to infer target gene expressions from clinically measured landmark expressions. Still, the huge number of genes to be inferred from a limited set of known expressions vexed the researchers. Ignoring possible correlation between target genes, they partitioned the target genes randomly and built separate networks to infer their expressions. This paper proposes that the dimensionality of the target set can be virtually reduced using deep autoencoders. Feedforward networks will be used to predict the coded representation of target expressions. In spite of the reconstruction error of the autoencoder, overall prediction error on the microarray based Gene Expression Omnibus (GEO) dataset was reduced by 6.6%, compared to D-GEX. An improvement of 16.64% was obtained on cross platform normalized data obtained by combining the GEO dataset and an RNA-Seq based 1000G dataset.