scGADR: Dimensionality Reduction of Single-Cell RNA-seq Data with ZINB-Based Graph Attention Autoencoder

Yongxuan Tang, Jiawei Luo, Zhongyuan Xu, Zehao Xiong, Bo Wang, Nguyen Hoang Tu

Abstract: Due to the high variability, high sparsity, and high dimensionality of single-cell RNA sequencing (scRNA-seq) data, dimensionality reduction is critical for visualizing and interpreting the high-dimensional scRNA-seq data. Herein, we propose a new single-cell graph attention dimensionality reduction method (scGADR) that utilizes graph attention autoencoders to reduce the dimensionality of the input data. scGADR employs the graph attention network to assign different weights to the cell’s neighbor nodes according to their importance and incorporates the zero-inflated negative binomial (ZINB) distribution into the model’s decoder to learn the low-dimensional latent representation of the original data. Additionally, the Kullback–Leibler (KL) algorithm facilitates self-optimizing iterative training. The effectiveness of scGADR is validated through downstream experiments on low-dimensional representation, demonstrating its superiority over six state-of-the-art dimensionality reduction methods on 13 representative datasets generated by different sequencing platforms.