Estimating Drivers of Cell State Transitions Using Gene Regulatory Network Models

Daniel Schlauch, Kimberly Glass, Craig P. Hersh, Edwin K. Silverman, John QuackenbushbioRxiv 2016 · DOI

Abstract

Specific cellular states are often associated with distinct gene expression patterns. These states are plastic, changing during development, or in the transition from health to disease. One relatively simple extension of this concept is to recognize that we can classify different cell-types by their active gene regulatory networks and that, consequently, transitions between cellular states can be modeled by changes in these underlying regulatory networks. Here we describe MONSTER, MOdeling Network State Transitions from Expression and Regulatory data, a regression-based method for inferring transcription factor drivers of cell state conditions at the gene regulatory network level. As a demonstration, we apply MONSTER to four different studies of chronic obstructive pulmonary disease to identify transcription factors that alter the network structure as the cell state progresses toward the disease-state. Our results demonstrate that MONSTER can find strong regulatory signals that persist across studies and tissues of the same disease and that are not detectable using conventional analysis methods based on differential expression.

Introduction

In the field of genomics, a cell's state—such as the distinction between "healthy" and "diseased"—is often treated as a static snapshot of gene activity. However, biological systems are fundamentally dynamic, governed by complex "wiring diagrams" known as gene regulatory networks (GRNs). When a cell transitions between states, it is not merely the abundance of specific gene products that changes; it is the underlying logic of the network itself.

Traditional genomic analyses rely heavily on Differential Expression (DE) to identify genes that are significantly more or less active in one state compared to another. While DE is a powerful tool, it often functions as a "black box." It identifies what changed but frequently fails to explain why. Transcription factors (TFs) act as the master switches of the cell, yet they do not always change their own expression levels to exert influence. Instead, they may alter which genes they target or the strength of their regulatory bonds. To capture these shifting relationships, we developed MONSTER (MOdeling Network State Transitions from Expression and Regulatory data).

The MONSTER Framework

MONSTER is a regression-based framework designed to identify the transcription factors driving cell state transitions by modeling changes in GRN structure. By treating a biological transition as a mathematical transformation between two networks, the method uncovers regulatory drivers that are often invisible to traditional analysis.

The MONSTER workflow operates in three primary stages:

  1. Network Inference: The method estimates state-specific GRNs (e.g., one for a control group and one for a disease group). This process utilizes a regression-based approach that integrates sequence motif data—representing the "prior" potential for regulation—with gene co-expression data, which reflects actual regulatory activity in the given samples.
  2. The Transition Matrix: This is the core innovation of the method. MONSTER models the transition between the two networks as a linear transformation: $B = AT + E$. In this equation, $A$ represents the initial network, $B$ represents the final network, and $T$ is the transition matrix.
  3. dTFI Calculation: By analyzing the "off-diagonal mass" of the transition matrix $T$, we calculate the differential Transcription Factor Involvement (dTFI). A high dTFI score indicates that a specific transcription factor has significantly altered its targeting patterns between the two states, suggesting it is a primary driver of the transition.

To ensure results are biologically meaningful rather than stochastic noise, MONSTER employs a permutation testing framework to generate a null distribution, allowing for the assignment of statistical significance to each TF's involvement.

Validation and Application in COPD

We validated MONSTER using in silico benchmarks and datasets from the DREAM5 challenge (including E. coli and S. cerevisiae). Following validation, we applied the method to four independent studies of Chronic Obstructive Pulmonary Disease (COPD) to test its performance in a complex disease context:

  • ECLIPSE and COPDGene: Focused on blood/PBMC samples.
  • LGRC and LT-CDNM: Focused on lung tissue samples.

These datasets provided a unique opportunity to determine if MONSTER could identify consistent regulatory signals across different cohorts and, more importantly, across different biological tissues (blood vs. lung).

Key Insights

The "Invisible" Driver

One of the most significant findings was that many of the top-ranked TFs identified by MONSTER—such as E2F4 and NRF1—showed no significant change in their own expression levels. In a standard DE analysis, these factors would have been overlooked. MONSTER reveals that their influence comes from a redistribution of their regulatory targets rather than a change in their abundance.

Network Stability and Plasticity

The MONSTER model assumes that the majority of a network remains stable during a transition (represented by the identity matrix). By specifically hunting for the "deviations" from this stability, the method isolates the specific regulatory changes that characterize the disease state.

Cross-Tissue Consistency

While DE results often vary significantly between blood and lung tissue, MONSTER’s dTFI scores were remarkably consistent ($r \approx 0.7-0.8$). This suggests that the underlying regulatory logic of a disease may be more conserved across different tissues than the expression of individual genes. This consistency highlights MONSTER's potential for identifying systemic biomarkers of disease.

Results and Conclusion

In the DREAM5 benchmarks, MONSTER significantly outperformed simple edge-weight subtraction methods in identifying true regulatory changes. In the COPD application, the method identified a robust set of TFs—including SP1, SP2, GABPA, and ELK1—as key drivers. These factors are biologically linked to cell cycle regulation and inflammation, both of which are hallmarks of COPD progression.

While the current model assumes a linear transition ($B=AT$), which may simplify some non-linear biological interactions, it provides a powerful and statistically rigorous way to move beyond gene lists. By focusing on the transition of the network itself, MONSTER offers a clearer view of the mechanisms driving cellular change, providing a more robust foundation for understanding disease etiology and identifying therapeutic targets.

Summary updated

Abstract

Abstract Specific cellular states are often associated with distinct gene expression patterns. These states are plastic, changing during development, or in the transition from health to disease. One relatively simple extension of this concept is to recognize that we can classify different cell-types by their active gene regulatory networks and that, consequently, transitions between cellular states can be modeled by changes in these underlying regulatory networks. Here we describe MONSTER , MO deling N etwork S tate T ransitions from E xpression and R egulatory data, a regression-based method for inferring transcription factor drivers of cell state conditions at the gene regulatory network level. As a demonstration, we apply MONSTER to four different studies of chronic obstructive pulmonary disease to identify transcription factors that alter the network structure as the cell state progresses toward the disease-state. Our results demonstrate that MONSTER can find strong regulatory signals that persist across studies and tissues of the same disease and that are not detectable using conventional analysis methods based on differential expression. An R package implementing MONSTER is available at github.com/QuackenbushLab/MONSTER.
BibTeX
@misc{schlauch2016estimating,
  author = {Daniel Schlauch and Kimberly Glass and Craig P. Hersh and Edwin K. Silverman and John Quackenbush},
  title = {{Estimating Drivers of Cell State Transitions Using Gene Regulatory Network Models}},
  year = {2016},
  doi = {10.1101/089003},
  howpublished = {bioRxiv},
}
Figure