ScDA: A Denoising AutoEncoder Based Dimensionality Reduction for Single-cell RNA-seq Data

Abstract

Single-cell RNA-seq (scRNA-seq) data has provided a higher resolution of cellular heterogeneity. However, scRNA-seq data also brings some computational challenges for its high-dimension, high-noise, and high-sparseness. The dimension reduction is a crucial way to denoise and greatly reduce the computational complexity by representing the original data in a low-dimensional space. In this study, to achieve an accurate low-dimension representation, we proposed a denoising AutoEncoder based dimensionality reduction method for scRNA-seq data (ScDA), combining the denoising function with the AutoEncoder. ScDA is a deep unsupervised generative model, which models the dropout events and denoises the scRNA-seq data. Meanwhile, ScDA can reveal the nonlinear feature extraction of the original data through maximum distribution similarity before and after dimensionality reduction. Tested on 16 scRNA-seq datasets, ScDA provides superior average performances, and especially superior performances in large-scale datasets compared with 3 clustering methods.