Quantifying the clusterness and trajectoriness of single-cell RNA-seq data

Hong Seo Lim, Peng QiuPLOS Computational Biology 2024 · DOI

TL;DR

Single-cell RNA sequencing (scRNA-seq) data can be interpreted as either distinct cell clusters or continuous developmental trajectories. This paper introduces a novel framework with five scoring metrics to objectively quantify these geometric properties, helping researchers choose the most appropriate analysis method for their data.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study biological systems at an unprecedented resolution. However, interpreting the complex data generated by these technologies presents a challenge. Two primary analytical approaches are commonly used: clustering, which identifies distinct cell types, and trajectory inference, which maps continuous cellular processes like differentiation or development. The problem arises because applying both methods to the same dataset can sometimes yield vastly different, even contradictory, interpretations. This ambiguity makes it difficult for researchers to confidently select the most suitable analysis method, potentially leading to misinterpretations of biological phenomena. Currently, the choice of analysis is often based on a researcher's prior biological intuition rather than objective data characteristics.

Methods

The core contribution of this paper is the development of five novel scoring metrics designed to quantify the "clusterness" and "trajectoriness" of scRNA-seq data. These metrics are based on different geometric properties of the data:

  • Pairwise Distance Distribution: Analyzes the entropy of distances between all cell pairs. Cluster-like data is expected to have a more "spiky" (multi-modal) distance distribution, while trajectory-like data will have a smoother, more spread-out distribution.
  • Persistent Homology: Leverages topological data analysis to study the "birth" and "death" of connected components as a distance threshold increases. The entropy of these merge thresholds is used as a score.
  • Vector Magnitude: Calculates the magnitude of a summed vector derived from connecting cluster centers identified by K-means. Trajectory-like data is expected to have a stronger directional component, leading to a larger vector magnitude.
  • Ripley's K Function: Adapts a spatial statistics tool to quantify the difference between the data's spatial distribution and a uniform distribution. This metric is hypothesized to be higher for cluster-like data.
  • Degrees of Connectivity: Measures how easily cells can reach each other through a network of nearest neighbors. Trajectory-like data is expected to show higher connectivity even with fewer neighbors.

These five scores are then used to create a "geometric landscape" by projecting them onto a 2D UMAP space. This landscape serves as a reference to classify new datasets.

Datasets

The authors utilized two main types of datasets:

  • Simulated Datasets: A large set of 12,000 simulated 2D datasets were generated, comprising four types: clear clusters, clear trajectories, noisy clusters, and noisy trajectories. These were crucial for testing and validating the scoring metrics.
  • Real scRNA-seq Datasets: 169 publicly available scRNA-seq datasets from previous benchmark studies were used to evaluate the applicability of the proposed geometric landscape to real-world data. These datasets were chosen because they were previously categorized by researchers based on presumed geometric properties (e.g., "clusters," "organs," "tree," "linear").

Key Insights and Ideas

  • Objective Quantification is Possible: The paper demonstrates that the geometric properties of scRNA-seq data, specifically its "clusterness" and "trajectoriness," can be objectively quantified using a suite of computational metrics.
  • A Unified Geometric Landscape: By projecting the scores onto a UMAP space, the authors create a visual "map" where different geometries (cluster-like vs. trajectory-like) occupy distinct regions, providing an intuitive way to understand data structure.
  • Biological Intuition Can Be Misleading: The study highlights that in a significant portion of benchmarked datasets (around 30%), the presumed geometric intuition of researchers did not align with the quantitative predictions from the scoring metrics. This suggests that relying solely on biological priors might be insufficient or even misleading.
  • Clustering Can Reduce Trajectoriness: When individual clusters from a larger dataset are analyzed separately, their inherent trajectoriness can become less prominent, as they are isolated from their original continuous context.

Key Results

  • The five proposed scoring metrics were shown to effectively differentiate between simulated cluster-like and trajectory-like datasets, with combinations of scores providing enhanced classification power.
  • The UMAP projection of scores from 12,000 simulated datasets created a clear geometric landscape where clear clusters and clear trajectories were well-separated.
  • When applied to 169 real scRNA-seq datasets, the proposed scoring metrics and geometric landscape showed comparable distributions to the simulated data, indicating their general applicability.
  • The quantitative predictions for the geometric nature of the 169 real datasets agreed with the presumed geometry based on biological intuition in approximately 70% of cases.

Limitations and Future Work

  • Metric Completeness: The authors acknowledge that the five presented metrics might not be exhaustive and suggest that developing additional metrics could further enrich the captured geometric variations.
  • Complex Geometries: Datasets with more complex geometries, such as "clusters of trajectories" or "trajectories of clusters," might not be robustly classified by the current framework and may require more nuanced approaches, potentially involving analyzing both the original dataset and its individual clusters.
  • Future Directions: The authors propose that this framework can serve as a valuable tool for guiding the choice of analysis methods (clustering vs. trajectory inference) for scRNA-seq data. They also suggest that the quantitative assessment of data geometry is an important area that warrants further discussion and development within the single-cell research community.

Summary updated

BibTeX
@article{lim2024quantifying,
  author = {Hong Seo Lim and Peng Qiu},
  title = {{Quantifying the clusterness and trajectoriness of single-cell RNA-seq data}},
  journal = {PLOS Computational Biology},
  year = {2024},
  doi = {10.1371/journal.pcbi.1011866},
}
Figure