Quantifying Waddington’s epigenetic landscape: a comparison of single-cell potency measures

Jifan Shi, Andrew E. Teschendorff, Weiyan Chen, Luonan Chen, Tiejun LibioRxiv 2018 · DOI

TL;DR

This paper introduces a novel method, Markov-Chain Entropy (MCE), to quantify single-cell potency by modeling cellular differentiation as a Markov process on a gene interaction network. MCE proves more robust than existing methods in distinguishing between cell states of varying potency, offering a more accurate way to construct Waddington's epigenetic landscape.

What Problem Does This Paper Address?

For over 60 years, Waddington's epigenetic landscape model has been a powerful conceptual tool for understanding cellular differentiation. This model visualizes cell fate transitions as a ball rolling down a landscape, with stable cell states (like stem cells) residing in "basins" or "attractor states." A key aspect of this landscape is "potency," which represents a cell's potential to differentiate into various cell types. High potency cells, like embryonic stem cells, are at the top of the landscape, while terminally differentiated cells are at the bottom.

However, translating this conceptual model into a quantitative, biophysical framework has been challenging. Specifically, accurately measuring single-cell potency from gene expression data has been a persistent problem. Existing methods often struggle with robustness and can be influenced by factors like cell cycle. This paper aims to address this by developing a more reliable and interpretable method for quantifying single-cell potency, thereby enabling a more precise reconstruction of Waddington's landscapes.

Methods

The core of this paper is the introduction of Markov-Chain Entropy (MCE), a novel single-cell potency measure. MCE is derived by:

  • Modeling cellular differentiation as a Markov process: The authors assume that a cell's gene expression profile represents the steady-state outcome of an underlying signaling process.
  • Integrating gene expression with a protein-protein interaction (PPI) network: This network provides a structural constraint on the possible signaling pathways within a cell.
  • Estimating the transition matrix: The method optimizes a Markov transition matrix that best explains the observed gene expression profile and adheres to the network's topology.
  • Calculating graph entropy: The MCE is then calculated as the graph entropy of this optimized Markov process. This entropy is interpreted as a measure of cell potency, with higher entropy indicating higher potency (e.g., pluripotency).

The authors also compare MCE to three existing potency measures: StemID, SLICE, and SCENT.

Datasets

The study utilizes a comprehensive collection of single-cell RNA-Sequencing (scRNA-Seq) datasets for evaluation:

  • 9 independent scRNA-Seq experiments: These experiments encompass a total of approximately 6,600 high-quality single-cell profiles.
  • 11 independent comparisons: These comparisons were devised from the scRNA-Seq data to specifically test the ability of potency measures to discriminate between cell types expected to have different potencies.
  • 2 studies with bulk RNA-Seq data: These were used to assess the robustness of the potency measures on aggregated cell populations.

The paper provides detailed descriptions of each dataset used in the "Online Methods" section, including information on cell types, experimental conditions, and data accession numbers. No new benchmark datasets were introduced; instead, existing datasets were used to evaluate the proposed method against established ones.

Key Insights and Ideas

  • MCE as a robust measure of potency: The central finding is that MCE provides a more stable and reliable estimate of single-cell potency compared to existing methods.
  • Network integration is crucial: The robustness of MCE (and SCENT) stems from their integration of gene expression data with a network structure, suggesting that cellular potency is encoded not just in gene expression levels but also in the network's connectivity patterns.
  • Hub genes and potency: The authors' analysis suggests that genes with higher connectivity in the network (hubs) are particularly important for determining cell potency. Cells expressing higher levels of these hub genes tend to have higher potency.
  • MCE enables accurate Waddington landscape reconstruction: The paper demonstrates that MCE can be used to construct visually informative and biologically relevant Waddington-like epigenetic landscapes, providing a quantitative basis for this conceptual model.
  • Cell-cycle adjustment is important but challenging: While adjusting for cell-cycle phase can improve potency estimates, it's a complex issue that requires careful consideration, as simple linear adjustments are often inappropriate.

Key Results

  • Superior discrimination accuracy: In 9 out of 11 comparisons, MCE and SCENT outperformed StemID and SLICE in discriminating between cell types with different potencies. In all 11 comparisons, MCE and SCENT were better than SLICE.
  • Robustness on bulk data: MCE and SCENT also showed robust performance on bulk RNA-Seq data, indicating their ability to capture essential potency information even when cell-level resolution is lost.
  • Accurate differentiation trajectory modeling: MCE successfully predicted the gradual decrease in potency during differentiation in several time-course experiments, including the differentiation of hESCs into neurons and mouse hepatoblasts into hepatocytes and cholangiocytes.
  • Network integration improves discrimination: Figure 4 demonstrates that integrating network information significantly improves the discrimination accuracy of MCE between neural progenitor cells and neurons, highlighting the importance of the connectome.
  • Quantitative landscape construction: Figure 1B showcases a 3D representation of a Waddington landscape constructed using MCE, illustrating how different cell types (hESCs, NPCs, etc.) occupy distinct attractor states.

Limitations and Future Work

  • Limitations of PPI networks: The authors acknowledge that PPI networks are simplifications of complex cellular signaling. Future work could explore using more refined gene regulatory networks (GRNs).
  • Computational complexity: While the paper presents an iterative method to solve the MCE optimization problem, it can still be computationally intensive for very large networks.
  • Cell-cycle confounding: The paper highlights the difficulty of fully disentangling cell-cycle effects from potency, especially in certain cell types like hESCs. Further research is needed to develop more robust methods for this.
  • Future directions: The authors propose that future single-cell analysis algorithms should focus more on developing and testing improved in-silico models of cell potency, as this is a key step towards more accurate and unbiased reconstruction of Waddington landscapes. They suggest that MCE provides a strong foundation for such advancements.

Summary updated

Abstract

Over 60 years ago Waddington proposed an epigenetic landscape model of cellular differentiation, whereby cell-fate transitions are modelled as canalization events, with stable cell states occupying the basins or attractor states 1, 2 . A key ingredient of this landscape is the energy potential, or height 3 , which correlates with cell-potency. To date, very few explicit biophysical models for estimating single-cell potency have been proposed. Using 9 independent experiments, encompassing over 6,600 high-quality single-cell RNA-Seq profiles, we here demonstrate that single-cell potency can be approximated as the graph entropy of a Markov Chain process on a model signaling network. Our analysis highlights that other proposed single-cell potency measures are not robust, whilst also revealing that integration with orthogonal systems-level information improves potency estimates. Thus, this study provides a foundation for an improved systems-level understanding of single-cell potency, which may have profound implications for the discovery of novel stem-and progenitor cell phenotypes.
BibTeX
@misc{shi2018quantifying,
  author = {Jifan Shi and Andrew E. Teschendorff and Weiyan Chen and Luonan Chen and Tiejun Li},
  title = {{Quantifying Waddington’s epigenetic landscape: a comparison of single-cell potency measures}},
  year = {2018},
  doi = {10.1101/257220},
  howpublished = {bioRxiv},
}

Loading related content...

Figure