Information retrieval in single cell chromatin analysis using TF-IDF transformation methods

TL;DR

This paper evaluates various transformation and dimension reduction methods, including TF-IDF, SVD, and autoencoders, to improve the analysis of high-dimensional, sparse scATAC-seq data for better cell clustering and feature extraction.

Contribution

It provides a comprehensive comparison of transformation techniques for scATAC-seq data, highlighting the effectiveness of TF-IDF and its impact on downstream analysis.

Findings

TF-IDF improves clustering accuracy

TF-IDF enhances biologically relevant feature extraction

Autoencoders benefit from TF-IDF transformation

Abstract

Single-cell sequencing assay for transposase-accessible chromatin (scATAC-seq) assesses genome-wide chromatin accessibility in thousands of cells to reveal regulatory landscapes in high resolutions. However, the analysis presents challenges due to the high dimensionality and sparsity of the data. Several methods have been developed, including transformation techniques of term-frequency inverse-document frequency (TF-IDF), dimension reduction methods such as singular value decomposition (SVD), factor analysis, and autoencoders. Yet, a comprehensive study on the mentioned methods has not been fully performed. It is not clear what is the best practice when analyzing scATAC-seq data. We compared several scenarios for transformation and dimension reduction as well as the SVD-based feature analysis to investigate potential enhancements in scATAC-seq information retrieval. Additionally, we investigate if autoencoders benefit from the TF-IDF transformation. Our results reveal that the TF-IDF transformation generally leads to improved clustering and biologically relevant feature extraction.