Gene Regulatory Network Inference from Gene Expression Dataset using Autoencoder

Genes, specific stretches of DNA sequences, encode for gene products to perform a number of biological functions required for life. Gene expression is the process of synthesizing gene products from the genetic code. Complex biological functions are controlled by a tight regulation of these interdependent gene expressions. High-level organisms with large genomes shows perplexing patterns of regulatory interactions between large number of genes, hence, inferring the gene regulatory network from small number of gene expression profiles is highly challenging task. In this study, we propose an autoencoder architecture to explore regulatory relationships among genes from a gene expression dataset consisting of a large number of experiments. Autoencoder is trained by a gene expression dataset prepared by NIH LINCS program which consist of 100,000 mRNA measurements for the 12,320 genes. We observed that Autoencoder is capable of reproducing gene expressions fairly well (average MSE 0.0045) with only 50 hidden states. We conjecture that learned autoencoder weights (input to hidden and hidden to output) can be used to predict regulatory interactions between genes. Furthermore, we will investigate hidden state representations to check whether they conform to biological states which dictates certain gene expression responses.