scMomentum: Inference of Cell-Type-Specific Regulatory Networks and Energy Landscapes

Larisa M. Soto, Juan P. Bernal-Tamayo, Robert Lehmann, Subash Balsamy, Xabier Martinez‐de‐Morentin, Amaia Vilas‐Zornoza, Patxi San-Martin, Felipe Prósper, David Gómez-Cabrero, Narsis A. Kiani, Jesper TegnérbioRxiv 2020 · DOI

TL;DR

This paper introduces scMomentum, a novel computational method that infers cell-type-specific gene regulatory networks (GRNs) and developmental energy landscapes directly from single-cell RNA sequencing (scRNA-seq) data. By integrating RNA velocity, scMomentum offers a powerful, efficient, and scalable approach to understanding cellular differentiation without needing time-series or perturbation experiments.

What Problem Does This Paper Address?

Understanding how cells differentiate and establish their unique identities is a fundamental question in biology. Gene regulatory networks (GRNs) are key to this process, dictating which genes are turned on or off. While single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study individual cells, inferring these complex GRNs from scRNA-seq data remains a significant challenge. Existing methods often struggle with scalability, computational efficiency, and the need for experimental data like time-series or perturbation experiments, which are not always available or practical. This paper aims to bridge this gap by providing a method that can robustly infer GRNs and visualize cellular developmental potential from static scRNA-seq data.

Methods

The core of scMomentum lies in its integration of RNA velocity, a concept that captures the dynamics of gene expression by considering both spliced and unspliced mRNA molecules. The method models the change in gene expression over time as a function of the expression of other genes and their degradation rates. This is formulated as a linear model where the gene-weighted and directed adjacency matrix represents the inferred GRN.

Key methodological contributions include:

  • RNA Velocity Integration: Leveraging scVelo to estimate RNA velocity and degradation rates, providing a dynamic view of gene expression.
  • Model-Based GRN Inference: A data-driven formulation to predict cell-type-specific GRNs by solving a system of equations based on RNA velocity.
  • Energy Landscape Reconstruction: Utilizing a Discrete Hopfield Network (DHN) formalism, where the inferred GRN acts as the interaction network. This allows for the reconstruction of a "Waddington-like" energy landscape, where cells with higher energy are considered to have greater developmental potential.
  • Computational Efficiency and Scalability: The linear model and formulation are designed to be computationally efficient and scalable to large datasets.

Datasets

The authors evaluated scMomentum using a combination of simulated and public datasets:

  • Simulated Dataset: 20,000 cells and 500 genes generated using Dyngen, with two independent branching trajectories.
  • Public Datasets:
    • Human fetal forebrain (hFB18)
    • Human PBMCs (hPB20)
    • Mouse developing spinal cord (mSC19)
    • Mouse brain atlas (mBA18)
    • Human embryonic hematopoiesis (hED19)
    • Human hematopoiesis - In house (mBD20)

They also benchmarked scMomentum against existing methods like GENIE3 and GRNBOOST2, using a novel metric based on the preservation of cluster distances (Mantel correlation between expression-derived and network-derived distance matrices).

Key Insights and Ideas

  • Unlocking Developmental Dynamics from Static Data: The most significant insight is the ability to infer dynamic processes like cell differentiation and developmental potential from seemingly static scRNA-seq data by cleverly incorporating RNA velocity.
  • Bridging GRNs and Developmental Landscapes: scMomentum elegantly connects the inference of gene regulatory networks to the visualization of cellular developmental potential, providing a more holistic understanding of cell fate decisions.
  • "Momentum" in Cell Fate: The name "scMomentum" aptly reflects the method's ability to capture the "momentum" or directionality of cells as they move through developmental stages, visualized as flow on the energy landscape.
  • Robustness to Noise: The method demonstrates robustness to cellular noise and information loss, a crucial feature for real-world biological data.

Key Results

  • Accurate GRN Inference: scMomentum successfully inferred cell-type-specific GRNs that preserved biological relationships, as evidenced by high Mantel correlations with expression-derived distances.
  • Computational Advantage: scMomentum significantly outperforms existing methods in terms of computational efficiency and scalability, inferring GRNs for ~100K cells in about a minute, while other methods took much longer.
  • Reconstruction of Developmental Trajectories: The method accurately reconstructed developmental trajectories and energy landscapes for both simulated and real biological systems, such as hematopoiesis and lymphoid lineages.
  • Identification of Key Regulators: scMomentum identified essential network regulators, a common feature of biological networks, which was not consistently observed with other methods.
  • Biological Insights: Analysis of the Prothymosin Alpha (PTMA) gene in hematopoiesis revealed interesting insights into its dynamic regulation, suggesting that its expression changes might be a consequence of altered cellular development rather than the cause.

Limitations and Future Work

  • Assumption of Linear Dynamics: The current formulation relies on a linear approximation of regulatory signals within quasi-stable attractors. While this ensures efficiency, it might not fully capture highly non-linear dynamics.
  • Data Requirements: The method's effectiveness is dependent on the quality and coverage of the scRNA-seq data, particularly for capturing RNA velocity.
  • Future Directions: The authors suggest that scMomentum's scalability and ability to infer directed, signed, and weighted GRNs open doors for building comprehensive Cell Regulatory Atlases and studying regulatory processes across diverse biological contexts. Further exploration of its application to other dynamic biological processes is also a promising avenue.

Summary updated

Abstract

Abstract Recent progress in single-cell genomics has generated multiple tools for cell clustering, annotation, and trajectory inference; yet, inferring their associated regulatory mechanisms is unresolved. Here we present scMomentum, a model-based data-driven formulation to predict gene regulatory networks and energy landscapes from single-cell transcriptomic data without requiring temporal or perturbation experiments. scMomentum provides significant advantages over existing methods with respect to computational efficiency, scalability, network structure, and biological application. Availability scMomentum is available as a Python package at https://github.com/larisa-msoto/scMomentum.git
BibTeX
@misc{soto2020scmomentum,
  author = {Larisa M. Soto and Juan P. Bernal-Tamayo and Robert Lehmann and Subash Balsamy and Xabier Martinez‐de‐Morentin and Amaia Vilas‐Zornoza and Patxi San-Martin and Felipe Prósper and David Gómez-Cabrero and Narsis A. Kiani and Jesper Tegnér},
  title = {{scMomentum: Inference of Cell-Type-Specific Regulatory Networks and Energy Landscapes}},
  year = {2020},
  doi = {10.1101/2020.12.30.424887},
  howpublished = {bioRxiv},
}
Figure