Dimension Reduction and Clustering Models for Single-Cell RNA Sequencing Data: A Comparative Study

Chao Feng, Shufen Liu, Hao Zhang, Renchu Guan, Dan Li, Fengfeng Zhou, Yanchun Liang, Xiaoyue FengInternational Journal of Molecular Sciences 2020 · OpenReview

TL;DR

This paper tackles the challenge of analyzing high-dimensional, noisy single-cell RNA sequencing (scRNA-seq) data by comparing various dimension reduction and clustering techniques. They found that feature selection is crucial, and while feature extraction generally improves clustering, its effectiveness varies with the method and the resulting feature space size.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular heterogeneity, generating massive datasets. However, these datasets are notoriously high-dimensional (thousands of genes per cell), sparse (many genes have zero expression in a given cell), and noisy. These characteristics make traditional data analysis methods, particularly clustering for cell type identification, ineffective. The "curse of dimensionality" means that distances between cells become less meaningful in such high-dimensional spaces, hindering accurate grouping. This paper aims to provide a systematic evaluation of different approaches to overcome these challenges, specifically focusing on dimension reduction and clustering methods to establish a reliable analysis pipeline.

Methods

The study employs a comparative approach, evaluating combinations of dimension reduction and clustering algorithms.

Dimension Reduction Methods: * Feature Selection: Filter-based feature selection (FBFS) based on gene expression variance was used. This method selects genes with the highest variance, assuming they are more biologically significant for distinguishing cell types. * Feature Extraction: * Principal Component Analysis (PCA): A linear method that projects data into a lower-dimensional space while retaining maximum variance. * Independent Component Analysis (ICA): A method that aims to decompose a multivariate signal into statistically independent non-Gaussian components. * Non-negative Matrix Factorization (NMF): A matrix decomposition technique used for non-negative data.

Clustering Algorithms: The study evaluated five classical and widely used clustering algorithms, representing different categories:

  • Hierarchical Clustering: A bottom-up approach that builds a hierarchy of clusters.
  • K-means: A squared error-based algorithm that partitions data into k clusters.
  • Fuzzy C-means: A fuzzy clustering algorithm that assigns membership degrees to each cluster.
  • DBSCAN: A density-based algorithm capable of discovering clusters of arbitrary shapes.
  • Louvain: A graph-theory-based algorithm for community detection, known for its efficiency and accuracy.

The researchers tested these algorithms individually and in combination with the dimension reduction techniques.

Datasets

The study utilized two publicly available, large-scale scRNA-seq datasets:

  • Mouse Cortex Data (GSE60361): This dataset comprises 3005 high-quality single cells from the cerebral cortex of a mouse, with 19,972 genes measured. It was used for initial experiments and evaluations.
  • Mouse Visual Cortex Data (GSE71585): This dataset, used to validate findings, contains data from the visual cortex of adult mice and includes seven distinct cell types.

Key Insights and Ideas

  • Feature Selection is a Must: For high-dimensional and sparse scRNA-seq data, selecting informative genes (feature selection) significantly improves clustering performance. This step helps to reduce noise and the impact of the "curse of dimensionality."
  • Feature Extraction Can Boost Performance, But It's Not a Silver Bullet: While feature extraction generally enhances clustering accuracy, its effectiveness is highly dependent on the chosen method and the resulting dimensionality of the reduced feature space.
  • PCA is a Robust Workhorse: Principal Component Analysis (PCA) demonstrated consistent and steady performance across various scenarios, making it a reliable choice for feature extraction.
  • ICA Shines in Small Spaces, But Can Be Tricky: Independent Component Analysis (ICA) performed well when reducing data to very small feature spaces. However, its performance deteriorated significantly as the feature space size increased, suggesting it might be sensitive to the underlying dependencies in gene expression.
  • Louvain is Often the Champion: The Louvain clustering algorithm consistently delivered strong results, often outperforming other methods, especially when combined with appropriate preprocessing.
  • Not All Combinations Are Equal: The study highlights that certain combinations of dimension reduction and clustering algorithms are more effective than others. For instance, ICA was found to be unsuitable for Fuzzy C-means clustering.

Key Results

  • Feature Selection Improvement: Using FBFS with the top 500 genes on the mouse cortex data improved clustering accuracy by up to 28.22% for the Louvain algorithm compared to using all 19,972 genes.
  • Feature Extraction Gains: When reducing the mouse cortex data to 20 dimensions using NMF, hierarchical clustering saw a 34.69% accuracy increase. K-means combined with ICA achieved a remarkable 58.18% improvement.
  • PCA's Steadiness: PCA consistently provided good results and showed less performance degradation with increasing dimensions compared to ICA and NMF.
  • Louvain's Dominance: On the mouse cortex data, Louvain clustering achieved the highest accuracy (0.73) without feature extraction, and this improved to 0.77 with feature selection.
  • ICA's Small-Space Success: Louvain combined with ICA achieved the highest accuracy (around 0.93) on the mouse cortex data when the feature space was small (less than 100 dimensions).
  • Visual Cortex Data Consistency: Similar trends were observed on the mouse visual cortex dataset, with feature selection and extraction generally improving clustering performance. Fuzzy C-means performed best on a subset of this data.

Limitations and Future Work

  • ICA's Dimensionality Sensitivity: The paper notes that ICA's performance degrades in larger feature spaces, suggesting it might not be universally applicable.
  • Combination Specificity: The study found that certain combinations (e.g., ICA with Fuzzy C-means) performed poorly, indicating that algorithm pairing is critical.
  • Future Work: The authors suggest that further investigation into optimal combinations and the development of more robust feature extraction methods could lead to even better clustering results. They also imply that understanding the biological basis for gene dependencies could further refine these methods.

Summary updated

Abstract

With recent advances in single-cell RNA sequencing, enormous transcriptome datasets have been generated. These datasets have furthered our understanding of cellular heterogeneity and its underlying mechanisms in homogeneous populations. Single-cell RNA sequencing (scRNA-seq) data clustering can group cells belonging to the same cell type based on patterns embedded in gene expression. However, scRNA-seq data are high-dimensional, noisy, and sparse, owing to the limitation of existing scRNA-seq technologies. Traditional clustering methods are not effective and efficient for high-dimensional and sparse matrix computations. Therefore, several dimension reduction methods have been introduced. To validate a reliable and standard research routine, we conducted a comprehensive review and evaluation of four classical dimension reduction methods and five clustering models. Four experiments were progressively performed on two large scRNA-seq datasets using 20 models. Results showed that the feature selection method contributed positively to high-dimensional and sparse scRNA-seq data. Moreover, feature-extraction methods were able to promote clustering performance, although this was not eternally immutable. Independent component analysis (ICA) performed well in those small compressed feature spaces, whereas principal component analysis was steadier than all the other feature-extraction methods. In addition, ICA was not ideal for fuzzy C-means clustering in scRNA-seq data analysis. K-means clustering was combined with feature-extraction methods to achieve good results.
BibTeX
@article{feng2020dimension,
  author = {Chao Feng and Shufen Liu and Hao Zhang and Renchu Guan and Dan Li and Fengfeng Zhou and Yanchun Liang and Xiaoyue Feng},
  title = {{Dimension Reduction and Clustering Models for Single-Cell RNA Sequencing Data: A Comparative Study}},
  journal = {International Journal of Molecular Sciences},
  year = {2020},
  url = {https://openreview.net/forum?id=hAheYsN8Y9},
}
Figure