Variational Auto Encoder Approach To Find Deferentially Expressed Genes

Abstract

A study of differentially expressed genes across different cell types will help in identifying cell-specific responses to treatments or diseases. Recent advances in single-cell technology enable an analysis of thousands of cells which brought lots of computational challenges in terms of noise in the data sets and required computational power to handle the big data. In recent years it has been found that the deep learning model is being used as a biological model for single-cell analysis. Using state-of-the-art techniques in deep learning successfully extracts non-linear feature set from single-cell data and is used for various downstream analysis.

Recently, deep learning models such as Autoencoder (AE) and Variational Autoencoder (VAE) models are being used to capture hidden patterns from single-cell gene expression data. In this paper, I proposed a framework that is based on a variational autoencoder called BiDiffVAE (Bi-directional Differential Variational Autoencoder) to extract differently expressed genes. The proposed method makes use of cluster distribution on every latent space and merged weights in the decoder to assign genes to a cluster. My results discovered new sets of genes that were not shown using state-of-the-art techniques and can properly rank the top genes based on their significance in making clustering.