Alzheimer's scRNA-seq Data Analysis Using Multi-type Deep Autoencoders

Li Xu, You Wu, Xue Tong, LiYanli, liushuaipeng

Keywords: Deep learning; scRNA-seq; Alzheimer's disease; Autoencoder; Imputation

TL;DR: We use multiple encoders to analyze scRNA, promoting the analysis of Alzheimer's disease.

Abstract: Single-cell RNA sequencing (scRNA-seq) technology has been applied in Alzheimer's disease (AD) research to explore its pathogenic mechanism. The complexity of analyzing and utilizing sequencing data is significantly amplified by the high dimensionality and high noise levels of the data, as well as the presence of missing data. In order to tackle these problems, we suggest implementing a novel data processing framework that consists of two primary algorithms: the imputation algorithm scICLGAE and the clustering algorithm scCapsZB. scICLGAE employs two graph autoencoders (GAE) to fill in missing values by utilizing comparison learning to filter similar nodes from both global information and local structure. In order to verify the imputation impact and enhance the precision of the clustering outcomes, we have devised the scCapsZB algorithm. scCapsZB is a method that integrates a capsule network and a zero-inflated negative binomial distribution (ZINB) autoencoder. It incorporates prior knowledge through the capsule network and employs a self-attention routing mechanism to reduce the number of training parameters. Additionally, it uses the ZINB model to capture the feature representation of the data. The testing of our new framework on both generic and Alzheimer's datasets demonstrates substantial enhancements.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 23999