Coordinated gene expression variability encodes the regulatory state of cells

By combining machine learning, theoretical modeling, and single-cell RNA sequencing data, this study reveals that coordinated cell-wide variability in gene expression serves as a distinct signature of cellular identity and regulatory organization, particularly distinguishing stem and progenitor cells from heterogeneous populations.

The Gist

The Big Idea: It's Not Just About the Volume, It's About the Chorus

Imagine a city where every building (a cell) has a loudspeaker system playing a song (the genes).

For a long time, scientists thought that to understand a building, you just needed to listen to the average volume of the music. If a building is loud, it's a factory; if it's quiet, it's a library.

But this paper argues that the average volume is only half the story. The real secret lies in the noise and how the speakers fluctuate. Do the speakers in a building all stutter at the exact same time? Do they all get louder and quieter in a synchronized dance, or do they just buzz randomly?

The authors discovered that how cells "stutter" together tells us exactly what kind of cell it is and what it is capable of doing.

---

The Analogy: The Orchestra vs. The Crowd

To understand the difference between a Stem Cell (a versatile, "super-power" cell) and a Differentiated Cell (a specialized, "finished" cell), imagine two different musical groups:

1. The Stem Cell: The Improvisational Jazz Band

Think of a stem cell as a jazz band where every musician is listening intently to every other musician.

• The Vibe: If the drummer hits a rim-shot, the bassist immediately adjusts their rhythm, and the saxophone player changes their note to match.
• The Result: The music is highly coordinated. Even though there is noise and improvisation, the fluctuations are linked. They are "in sync."
• The Meaning: This high level of coordination means the cell is flexible. It can easily switch genres (change into a skin cell, a heart cell, etc.) because the whole network is talking to each other.

2. The Differentiated Cell: The Crowd at a Concert

Now, think of a specialized cell (like a skin cell) as a crowd of people at a concert.

• The Vibe: Everyone is doing their own thing. One person is clapping, another is cheering, another is checking their phone. Their movements are random and independent.
• The Result: The fluctuations are uncorrelated. If one person jumps, it doesn't mean the person next to them will jump.
• The Meaning: This cell is "set in its ways." It knows exactly what it is (a skin cell) and isn't likely to turn into something else. The noise is just random static, not a coordinated signal.

---

The Discovery: Listening to the "Static"

The researchers developed a new way to listen to this "static" in single-cell data (scRNA-seq). Usually, scientists try to filter out the noise to find the clear signal. These scientists said, "Wait, the noise is the signal!"

They created a mathematical score (called $\omega_{max}$) that measures how "together" the fluctuations are.

• High Score: The genes are fluctuating in a coordinated, synchronized way. This means the cell is likely a Stem Cell or a Progenitor Cell (a cell that is still deciding its fate).
• Low Score: The genes are fluctuating randomly and independently. This means the cell is Differentiated (specialized).

Why This Matters: The "Superpower" Detector

The paper shows that this "coordination score" is a superpower detector.

1. Finding Hidden Potential: They looked at cells that were supposed to be "finished" (differentiated). Some of them had a high coordination score. When they dug deeper, they found these cells were actually regenerating or healing. For example, in a damaged gut, some mature cells suddenly started acting like stem cells to fix the injury. The "noise" pattern gave it away before the cell even changed its appearance.
2. Detecting Disease: They looked at heart regeneration in fish and lung disease in humans. They found that when a tissue is injured or sick, the cells start "jazzing" again. Their genes start fluctuating in a coordinated way to try to fix the problem.
3. The "Ghost" in the Machine: Sometimes, a group of cells looks like one type, but the "noise" pattern reveals a hidden subgroup inside that is actually different. It's like hearing a choir where one section suddenly starts singing a different song; the math can spot that hidden group even if you can't see it yet.

The Bottom Line

This paper changes how we look at cells.

• Old View: A cell is defined by what genes are turned ON (the average volume).
• New View: A cell is also defined by how its genes WIGGLE TOGETHER (the coordinated noise).

Just as a jazz band's ability to improvise together shows they are ready for anything, a cell's ability to fluctuate in sync shows it has the potential to change, heal, or regenerate. By listening to the "chaos," we can understand the order.

Technical Summary

1. Problem Statement

Gene expression is inherently stochastic, leading to cell-to-cell variability in mRNA and protein abundances. While single-cell RNA sequencing (scRNA-seq) has revealed this variability, a major challenge remains in distinguishing between:

• Biological noise: Functional stochasticity that may facilitate stress responses or differentiation.
• Technical noise: Artifacts inherent to sequencing technologies.
• Population heterogeneity: Variability arising from mixing distinct cell types rather than intrinsic stochasticity within a single state.

Current metrics often focus on mean expression levels or transcriptional diversity (entropy) within individual cells. However, it is unclear whether coordinated fluctuations across the entire transcriptome (cell-wide coordination) exist and whether they encode specific regulatory states (e.g., stemness vs. differentiation) that are invisible to mean-based analyses.

2. Methodology

The authors developed a hybrid framework combining statistical physics, theoretical modeling, and machine learning to analyze scRNA-seq data.

A. Theoretical Framework (Statistical Physics)

• Minimal Model: The authors modeled gene regulatory networks (GRNs) as a system of $N$ genes where mRNA and protein levels fluctuate around steady states.
• Hamiltonian Formulation: They derived a joint probability distribution for fluctuations ($\epsilon$ for mRNA, $\vartheta$ for proteins) taking the form of a bipartite spin-glass model:
  $$H = \sum V(\epsilon) + \sum V(\vartheta) + \sum K \epsilon \vartheta + \sum J_{ij} \epsilon_i \vartheta_j$$
  Here, $J_{ij}$ represents effective gene-gene interaction strengths.
• Phase Transition Hypothesis: The theory predicts a phase transition between two regimes:
  1. Uncorrelated Regime: Fluctuations are independent; overlap distributions scale as $N^{-1/2}$ (Central Limit Theorem).
  2. Correlated Regime: Fluctuations are coordinated across the cell; overlap distributions show broad deviations from $N^{-1/2}$ scaling, indicating long-lived memory and collective dynamics.

B. Statistical Metrics

• Overlap ($Q$): Defined as the correlation of fluctuations between "replicas" (cells within the same annotated state). The width of the overlap distribution quantifies the degree of coordination.
• Scaling Analysis: The authors tested if the standard deviation of overlaps scales as $N^{-1/2}$ (uncorrelated) or slower (correlated) as the number of genes ($N$) increases.
• Leading Eigenvalue ($\omega_{max}$): They utilized the largest eigenvalue of the empirical covariance matrix of gene fluctuations as a model-free proxy for the strength of the dominant collective fluctuation mode.

C. Machine Learning (Restricted Boltzmann Machines - RBMs)

• Physics-Informed ML: The authors mapped their theoretical Hamiltonian to a Gaussian-Gaussian Restricted Boltzmann Machine (RBM).
  • Visible units ($v$) = mRNA fluctuations.
  • Hidden units ($h$) = Protein fluctuations.
  • Weights ($W_{ij}$) = Effective gene-gene interactions ($J_{ij}$).
• Inference: They trained RBMs on scRNA-seq data to infer the distribution of interaction strengths. They calculated a proxy parameter $\hat{\varpi} = N^{1/2} \cdot \text{SD}(W_{ij})$ to estimate the system's proximity to the phase transition.

D. Data Analysis Pipeline

• Datasets: Analyzed scRNA-seq data from:
  • Developing and adult mouse brain.
  • Human Pan-GI Cell Atlas (gastrointestinal tract).
  • Zebrafish heart regeneration.
  • Human pulmonary disease.
• Preprocessing: Technical noise was mitigated by focusing on scaling exponents (dimensionless) rather than absolute variance, and by using subsampling and k-nearest neighbor (kNN) controls to rule out cluster-size artifacts.

3. Key Contributions

1. Definition of Coordinated Variability: Introduced a statistical score ($\omega_{max}$ and overlap scaling) to quantify cell-wide coordination of gene expression noise, distinct from mean expression or single-cell entropy.
2. Theoretical-Experimental Bridge: Validated a spin-glass model of GRNs against real scRNA-seq data, demonstrating that biological systems operate near a critical phase transition.
3. RBM Framework for GRNs: Developed a physics-informed machine learning approach to infer effective gene-gene interaction structures from static single-cell data without prior knowledge of the network topology.
4. New Biomarker for Potency: Established that high coordination of fluctuations is a hallmark of stem and progenitor cells, while low coordination characterizes differentiated cells.

4. Key Results

• Stem vs. Differentiated States:
  • Stem/Progenitor Cells: Exhibit strongly correlated fluctuations. The overlap distribution width scales slower than $N^{-1/2}$, and the leading eigenvalue ($\omega_{max}$) is high. This suggests a "permissive" regulatory landscape with strong collective dynamics.
  • Differentiated Cells: Exhibit weakly correlated, independent fluctuations. The overlap scaling follows the Central Limit Theorem ($N^{-1/2}$), and $\omega_{max}$ is lower, indicating a "constrained" regulatory state.
• Dedifferentiation and Plasticity:
  • In the human gut, mature colonocytes (differentiated) that retain the capacity to dedifferentiate upon injury showed stem-like fluctuation patterns (high $\omega_{max}$ and high coordination).
  • This suggests that coordinated variability captures the latent potential of cells to change identity.
• Disease and Injury Detection:
  • Zebrafish Heart Regeneration: Injury-induced regenerative states (proliferative macrophages, regenerative fibroblasts) showed a peak in $\omega_{max}$, distinguishing them from homeostatic states without needing canonical marker genes.
  • Human Pulmonary Disease: The framework detected functional state changes in disease vs. control conditions based solely on covariance structure.
• Subpopulation Discovery:
  • The eigenvalue analysis revealed hidden subpopulations. For example, in the zebrafish data, a high $\omega_{max}$ in a "macrophage" cluster was traced to a small subpopulation of T-cells (contamination/misannotation) via eigenscore projection, demonstrating the method's ability to detect heterogeneity.

5. Significance

• Reframing Variability: The paper shifts the interpretation of gene expression noise from a nuisance or random error to a functional regulatory signature. It posits that the structure of variability (how fluctuations propagate) is as important as the mean expression level for defining cell identity.
• Robustness to Technical Noise: By relying on scaling laws and covariance structures rather than absolute counts, the method is robust to common scRNA-seq technical artifacts.
• Predictive Power: The framework allows for the inference of cellular potency and regulatory states (e.g., dedifferentiation potential, injury response) from static snapshots of data, even in the absence of known marker genes.
• Biophysical Insight: It provides a quantitative link between the microscopic dynamics of gene regulatory networks and macroscopic cell states, suggesting that stem cells operate near a critical point to maintain plasticity, while differentiation moves the system toward a more constrained, stable regime.

In summary, the authors demonstrate that coordinated gene-expression variability is a fundamental layer of cellular organization that encodes regulatory states, offering a powerful new tool for analyzing cell identity, plasticity, and disease states in single-cell genomics.