Analyzing scRNA-seq data by CCP-assisted UMAP and t-SNE

Yuta Hozumi, Gu-Wei WeiarXiv 2023 · DOI · arXiv

Abstract

Single-cell RNA sequencing (scRNA-seq) provides high-resolution insights into cellular diversity but presents significant computational challenges due to high dimensionality, sparsity, and noise. We present Correlated Clustering and Projection (CCP), a novel preprocessing method that enhances the performance of standard visualization tools like UMAP and t-SNE. By grouping genes into "super-genes" based on correlations rather than abstract mathematical transformations, CCP improves clustering accuracy by an average of 19% across multiple datasets.

Introduction

The analysis of scRNA-seq data typically relies on dimensionality reduction to visualize complex cellular landscapes. Standard pipelines often utilize Principal Component Analysis (PCA) followed by UMAP or t-SNE. However, these methods rely on matrix diagonalization, which can be computationally expensive and may fail to preserve global structures or biological interpretability. CCP addresses these issues by operating directly in the data domain, providing a more robust initialization for downstream visualization.

Methods

The CCP framework involves a two-step "data-domain" approach:

  1. Gene Clustering: Thousands of genes are partitioned into $N$ clusters based on their expression correlations across cells.
  2. Gene Projection: Within each cluster, genes are projected into a single "super-gene" using a Flexibility Rigidity Index (FRI) nonlinear kernel. These super-genes represent aggregated biological signals.

Unlike standard methods that discard low-variance genes, CCP groups them into a single descriptor, preserving collective signals that might otherwise be lost.

Results

The performance of CCP-assisted visualization was evaluated on eight GEO datasets, including human and mouse samples from embryonic, brain, and pancreatic tissues.

Improved Cluster Separation

A significant challenge in scRNA-seq visualization is the "hairball" effect, where distinct cell types overlap in the 2D projection. CCP-assisted t-SNE effectively resolves these overlaps, as demonstrated in the comparison below.

Comparison of CCP-assisted t-SNE vs. unassisted t-SNE, showing the resolution of overlapping clusters into distinct cell types.

Consistency Across Datasets

CCP-assisted methods consistently outperformed standard UMAP and t-SNE across diverse biological contexts, including embryonic development and immune cell populations. Quantitatively, the mean Adjusted Rand Index (ARI) for CCP-assisted UMAP was 0.752, compared to 0.630 for standard UMAP.

Performance comparison of CCP-assisted vs. unassisted UMAP and t-SNE across four representative scRNA-seq datasets.

The method was particularly effective at capturing continuous differentiation trajectories in time-series data (e.g., GSE75748), where standard methods often struggle to show clear boundaries between stages.

Discussion and Conclusion

CCP-assisted visualization provides a biologically grounded and robust framework for scRNA-seq analysis. By bypassing matrix diagonalization and leveraging gene-gene correlations, it offers superior cluster separation and stability. While future work is needed to optimize parameter selection and integrate the method into existing software suites like Seurat or Scanpy, CCP represents a significant step forward in high-dimensional biological data visualization.

Summary updated

BibTeX
@misc{hozumi2023analyzing,
  author = {Yuta Hozumi and Gu-Wei Wei},
  title = {{Analyzing scRNA-seq data by CCP-assisted UMAP and t-SNE}},
  year = {2023},
  eprint = {2306.13750},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.13750},
}
Figure