Learning Edge Rewiring in EMT From Single Cell Data

Smita Krishnaswamy, Nevena Zivanovic, Roshan Sharma, Dana Pe’er, Bernd BodenmillerbioRxiv 2017 · DOI

TL;DR

This paper introduces a novel computational framework to understand how cellular signaling networks dynamically rewire during transitions like Epithelial-to-Mesenchymal Transition (EMT), using static single-cell data. By leveraging pseudo-time trajectories and new metrics like TIDES and 3D-DREMI, they can predict and validate critical signaling events that drive these cellular state changes.

What Problem Does This Paper Address?

Cellular processes, such as development and disease progression, involve dynamic changes in cell states. These changes are driven by complex regulatory networks that are not static but constantly reconfigure. Traditional computational methods often analyze these networks from single time-point snapshots, treating them as static. This overlooks the crucial dynamic rewiring of interactions (edges) that occurs during transitions. The authors aim to bridge this gap by developing methods to infer these dynamic changes from static single-cell data, specifically focusing on the Epithelial-to-Mesenchymal Transition (EMT) as a model system. Understanding this rewiring is key to identifying critical events that drive these transitions and could inform therapeutic interventions.

Methods

The core of this work lies in developing novel computational methods to infer dynamic network rewiring from static single-cell data.

  • Pseudo-time Trajectory Inference: They use the Wanderlust algorithm to order cells along a one-dimensional trajectory that approximates the EMT progression. This "EMT-time" allows them to treat static snapshots as a continuous process.
  • 3D-DREVI (3D conditional Density Rescaled Visualization): This method extends existing 2D DREVI to visualize the relationship between two molecules (Y and Z) as a function of another molecule (X) and time (T, i.e., EMT-time). It helps characterize how relationships change across the trajectory.
  • 3D-DREMI (3D conditional Density Resampled Estimate of Mutual Information): This metric quantifies the strength of the relationship of a molecule Z on both time (T) and another molecule X, providing a score for how much information X and T together provide about Z.
  • TIDES (Trajectory Interpolated DREMI Scores): This is a key innovation. TIDES quantifies the relationship between two molecules continuously along EMT-time by computing 2D-DREMI on fixed-time slices of the 3D density estimate. This allows for tracking the strength of a specific edge's rewiring over the entire transition.

Datasets

The primary dataset used is mass cytometry data from Py2T murine breast cancer cells treated with TGF-beta for 2, 3, and 4 days. This dataset captures the expression of 32 markers, including epithelial and mesenchymal markers, signaling pathway components, and EMT transcription factors. The authors also generated data from acute and chronic drug perturbation experiments for validation.

Key Insights and Ideas

  • Static snapshots can reveal dynamic processes: The paper demonstrates that by cleverly ordering cells into a pseudo-time trajectory, static single-cell data can be used to infer dynamic changes in cellular networks.
  • EMT is a continuum, not a binary switch: The data clearly shows a continuous spectrum of cells transitioning from epithelial to mesenchymal states, highlighting the rate-heterogeneous nature of EMT.
  • Network rewiring is critical for EMT: The developed methods reveal that the strength of relationships between signaling molecules changes significantly during EMT, with specific "edges" becoming more or less important at different stages.
  • TIDES as a powerful tool for dynamic analysis: TIDES provides a quantitative and continuous view of how pairwise molecular relationships evolve, enabling the identification of critical points in the transition where information flow is high.
  • Predicting critical drivers of EMT: By identifying edges with high 3D-DREMI scores, the authors can pinpoint molecules and pathways that are likely critical drivers of EMT, which can then be experimentally validated.

Key Results

  • The Wanderlust algorithm successfully generated a one-dimensional EMT-time trajectory that recapitulated expected marker changes (e.g., decreasing E-cadherin, increasing Vimentin).
  • TIDES curves revealed dynamic changes in the strength of signaling relationships. For example, the pPLCy2-pMEK1/2 relationship peaked in strength during the transitional phase.
  • The authors identified several critical edges, including those involving pSMAD2/3, AMPK, and β-catenin, which were predicted to be important for EMT progression.
  • Validation with drug perturbations:
    • Inhibition of TGFβ-receptor reduced mesenchymal cells from 54% to 2%.
  • MEK inhibition reduced mesenchymal cells from 54% to 17%.
  • WNT inhibition reduced mesenchymal cells from 26% to 11%.
  • AMPK activation slightly increased EMT (52% to 60%).
  • AKT inhibition showed no significant impact, suggesting system-specific EMT drivers.
  • The developed methods (3D-DREVI, 3D-DREMI, TIDES) showed efficient runtime scaling with data size and number of bins.

Limitations and Future Work

  • Assumption of a similar path: A major assumption is that while cells progress at different rates, they largely follow a similar path. Deviations from this could affect trajectory inference.
  • Data sparsity: While the methods are designed for static snapshots, high-dimensional single-cell data can still be sparse, posing challenges for density estimation. The authors mention that their approach could be modified for sparse data conditions.
  • Extending to other data types: The authors suggest that their framework could be extended to other high-dimensional single-cell data types, such as single-cell RNA-sequencing, although this would require addressing data sparsity.
  • Causal interpretation: While DREMI suggests influence, a causal interpretation of edges might require prior biological knowledge or further experimental validation.
  • Future work could involve: applying these methods to understand regulatory systems governing malignant processes and assessing how perturbations affect highly rewired networks over entire continua of states.

Summary updated

Abstract

Abstract Cellular regulatory networks are not static, but continuously reconfigure in response to stimuli via alterations in gene expression and protein confirmations. However, typical computational approaches treat them as static interaction networks derived from a single experimental time point. Here, we provide a method for learning the dynamic modulation, or rewiring of pairwise relationships (edges) from a static single-cell data. We use the epithelial-to-mesenchymal transition (EMT) in murine breast cancer cells as a model system, and measure mass cytometry data three days after induction of the transition by TGFβ. We take advantage of transitional rate variability between cells in the data by deriving a pseudo-time EMT trajectory. Then we propose methods for visualizing and quantifying time-varying edge behavior over the trajectory and use these methods: TIDES (Trajectory Imputed DREMI scores), and measure of edge dynamism (3DDREMI) to predict and validate the effect of drug perturbations on EMT.
BibTeX
@misc{krishnaswamy2017learning,
  author = {Smita Krishnaswamy and Nevena Zivanovic and Roshan Sharma and Dana Pe’er and Bernd Bodenmiller},
  title = {{Learning Edge Rewiring in EMT From Single Cell Data}},
  year = {2017},
  doi = {10.1101/155028},
  howpublished = {bioRxiv},
}
Figure