Impact of data preprocessing on cell-type clustering based on single-cell RNA-seq data

Chunxiang Wang, Xin Gao, Juntao LiuBMC Bioinformatics 2020 · DOI

Abstract

Single-cell RNA sequencing (scRNA-seq) has revolutionized biological research by enabling cell-specific gene expression analysis. However, the accuracy of identifying cell types is highly dependent on data preprocessing. This study introduces SC3-e, an automated selector that identifies the optimal preprocessing transformation for the SC3 clustering algorithm. By utilizing a graph-based "C-score" metric, SC3-e significantly improves cell-type identification accuracy across diverse datasets.

Introduction

Single-cell RNA sequencing allows researchers to examine gene expression at an unprecedented resolution. Despite its potential, scRNA-seq data is notoriously noisy due to low capture efficiency and "dropouts." To identify distinct cell types, researchers rely on clustering algorithms, which are preceded by critical preprocessing steps such as log-transformation or normalization. Currently, there is no universal standard for preprocessing; different methods can lead to vastly different clustering outcomes depending on the dataset and algorithm used. This paper addresses this challenge by introducing SC3-e, a graph-based tool designed to automatically select the optimal preprocessing method for the widely used SC3 (Single-Cell Consensus Clustering) algorithm.

Methodology: The SC3-e Framework

The authors developed SC3-e to move beyond the "default" preprocessing trap. The framework operates through a multi-step pipeline:

  1. Multi-run Preprocessing: The input data is processed using four candidate methods: log transformation, z-score transformation, sctransform, and no transformation.
  2. Graph Construction and Contraction: For each method, a clustering graph is built based on SC3’s internal consensus matrix. This graph is then simplified into "cluster nodes."
  3. The C-score Metric: To evaluate the quality of the resulting clusters without knowing the ground truth, the authors introduced the C-score. This unsupervised metric calculates the ratio of within-cluster tightness ($W_{in}$) to between-cluster distinctness ($D_{out}$). A lower C-score indicates a more robust and accurate clustering result.
C-score performance across eight benchmark datasets. Red bars indicate the preprocessing method that achieved the maximum Adjusted Rand Index (ARI), demonstrating the C-score's predictive power.

Experimental Setup

The performance of SC3-e was evaluated using eight publicly available scRNA-seq datasets with known cell types: Darmanis, Lake, Yan, Romanov, Baron-5000, Biase, Deng, and Leng. These datasets represent a variety of normalization units (e.g., CPM, TPM, RPKM), providing a rigorous test for the algorithm's generalizability.

Key Results

The study yielded several significant findings:

  • Failure of Defaults: Interestingly, the default log transformation was not the optimal preprocessing method for any of the eight tested datasets.
  • Accuracy and Performance: SC3-e successfully identified the best preprocessing method in nearly every case. By selecting the optimal transformation, clustering accuracy (measured by the Adjusted Rand Index) improved by as much as 37% compared to default settings.
  • Stability: Because SC3 is stochastic, the authors tested the impact of repeating the C-score calculation. They found that averaging scores over just three runs ($N=3$) was sufficient to reach 100% selection accuracy for most datasets.
Selection accuracy as a function of the number of repeats (N) for calculating the C-score. High accuracy is achieved with minimal repetitions.

Discussion and Limitations

While SC3-e significantly enhances clustering performance, it has limitations. It is currently optimized specifically for the SC3 algorithm and may require adaptation for other popular tools like Seurat. Additionally, the requirement to run multiple preprocessing pipelines increases computational overhead. Finally, scalability remains a concern for "atlas-scale" datasets involving hundreds of thousands of cells, as the underlying SC3 algorithm has memory constraints. Future work will focus on adapting the C-score logic to more scalable clustering frameworks.

Summary updated

BibTeX
@article{wang2020impact,
  author = {Chunxiang Wang and Xin Gao and Juntao Liu},
  title = {{Impact of data preprocessing on cell-type clustering based on single-cell RNA-seq data}},
  journal = {BMC Bioinformatics},
  year = {2020},
  doi = {10.1186/s12859-020-03797-8},
}
Figure