Inferring division-associated stochasticity from time-series single-cell transcriptomes

The paper introduces scDIVIDE, a neural stochastic differential equation framework that leverages the coupling between birth rates and diffusion coefficients to accurately infer cell division dynamics and partitioning noise from time-series single-cell transcriptome data, thereby outperforming existing methods in predicting cell fate distributions.

The Gist

Imagine you are trying to figure out how a city grows and changes over time, but you can only take a few snapshots of the city at different moments. You see the buildings (cells) in the morning, afternoon, and evening, but you can't watch them move or multiply in real-time because the camera breaks the moment you try to look too closely.

This is the challenge scientists face with single-cell RNA sequencing (scRNA-seq). They can see what genes are active in thousands of individual cells at specific moments, but they can't track the same cell as it divides and changes.

The paper introduces a new tool called scDIVIDE (think of it as a "Time-Traveling Detective") that solves this mystery by understanding a specific rule of nature: when a cell divides, it creates chaos.

Here is the story of how scDIVIDE works, explained with simple analogies:

1. The Problem: The "Blurry Photo" of Cell Division

When a cell divides, it's like a photocopier making two copies of a document. But this photocopier is a bit glitchy. It doesn't make perfect copies; it adds a little bit of random noise (stochasticity) to each new copy. Some copies get a little more ink here, a little less there.

In the past, scientists tried to guess how fast a city (a population of cells) was growing just by looking at the total number of buildings. They assumed that if the population grew, the buildings just got bigger or more numerous. But they missed a crucial detail: the act of dividing itself creates a "blur" or "jitter" in the population.

If you have a quiet neighborhood where no one moves (quiescent cells), the houses stay still.
If you have a busy construction zone where houses are constantly being built and torn down (high turnover), the area looks chaotic and "jittery," even if the total number of houses stays the same.

Old methods couldn't tell the difference between a quiet neighborhood and a busy construction zone just by looking at the snapshots. They thought, "No net growth? Then nothing is happening."

2. The Solution: The "Glitch" is the Clue

The authors of this paper realized that this "glitch" (the noise from cell division) is actually a superpower.

They built a mathematical model based on a concept called Birth-Death-Mutation.

• Birth: A cell splits into two.
• Death: A cell disappears.
• Mutation (The Glitch): The two new cells are slightly different from the parent because the parts were shuffled randomly during the split.

The big discovery? The amount of "glitch" or chaos in the population is directly tied to how often cells are dividing.

Think of it like a dance floor:

• If people are just standing still, the crowd is smooth.
• If people are dancing wildly but staying in the same spot (high turnover), the crowd looks very blurry and chaotic.
• scDIVIDE looks at how "blurry" the crowd is and says, "Ah! This isn't just random noise; this blur means people are dancing (dividing) really fast!"

3. How scDIVIDE Works: The "Smart Simulator"

scDIVIDE is a computer program (a neural network) that acts like a super-smart simulator.

1. It takes the snapshots: It looks at the gene data from the morning, afternoon, and evening.
2. It runs a simulation: It imagines thousands of virtual cells moving and dividing.
3. It checks the "Blur": It asks, "If I make these cells divide at this speed, does the resulting 'blur' in my simulation match the real blurry photos we took?"
4. It learns: If the simulation is too smooth, it knows the cells aren't dividing enough. If it's too chaotic, they are dividing too fast. It adjusts its internal "birth rate" dial until the simulation matches reality perfectly.

Because it links the speed of division directly to the amount of chaos, it can figure out the birth rate without needing to count cells or assume the population size stays the same. It uses the physics of division as a built-in rule.

4. The Real-World Test: The Blood Factory

To prove it worked, the team tested scDIVIDE on data from mouse blood stem cells (the cells that make all your blood).

• The Mystery: In a lab dish, these stem cells were supposed to be growing. But standard tools were confused. Some said the immature cells were growing fast; others said they were slow.
• The scDIVIDE Answer: By looking at the "chaos" (noise), scDIVIDE correctly identified that the immature stem cells were actually dividing slowly, while the more mature cells were dividing faster.
• Why it matters: This matched what biologists had suspected from other experiments but couldn't prove with gene data alone. It also found that the genes controlling this "division speed" were related to chemical signals (cytokines) that tell cells when to multiply or die.

The Bottom Line

Before this paper, trying to guess how fast cells were dividing from gene data was like trying to guess how fast a car was driving just by looking at a blurry photo of its headlights. You couldn't be sure if the blur was from speed or just a bad camera.

scDIVIDE realized that the blur is the speedometer. By understanding that cell division creates specific types of noise, this new tool can accurately reconstruct the entire life story of a cell population—how they move, how they change, and most importantly, how fast they are reproducing—just from a few snapshots in time.

It turns the "noise" of life into a clear signal.

Technical Summary

1. Problem Statement

Cell division is a fundamental process in multicellular organisms, yet the impact of partitioning noise (stochastic fluctuations in cellular components during division) on genome-wide gene expression dynamics remains poorly understood.

• The Challenge: Traditional single-cell RNA sequencing (scRNA-seq) provides only discrete "snapshot" data, lacking continuous individual cell trajectories.
• The Limitation of Existing Methods: Current trajectory inference methods often treat cell division merely as a net population growth factor (birth minus death). They fail to account for the fact that division is a discrete stochastic event that introduces state perturbations (noise) into daughter cells. Consequently, existing models struggle to distinguish between:
  1. Quiescent systems: Low division activity, low noise.
  2. Homeostatic systems: High turnover (high birth and death rates) but near-zero net growth, which generates high noise due to frequent division events.
• The Gap: There is a lack of frameworks that can jointly infer continuous cellular dynamics, growth rates, and division-associated stochasticity from snapshot scRNA-seq data using biophysically grounded constraints.

2. Methodology: scDIVIDE

The authors propose scDIVIDE (single-cell DIVision-linked Inference of stochastic Dynamics in Expression states), a deep learning framework based on Neural Stochastic Differential Equations (SDEs).

Theoretical Foundation

• Birth-Death-Mutation (BDM) Processes: The model derives macroscopic population dynamics from microscopic BDM processes. It posits that cell division introduces a "mutation" term (partitioning noise) where daughter cells are displaced from the mother's state by a random vector.
• Fokker-Planck Equation (FPE): By taking the small-noise limit of the BDM process, the authors derive a macroscopic FPE for cell density $\xi(x,t)$:
  $$\frac{\partial \xi}{\partial t} = -\nabla \cdot (\mathbf{v}\xi) + g\xi + \frac{1}{2}\Delta [(\sigma^2 + 2\delta^2 b)\xi]$$
  • $\mathbf{v}$: Velocity field (differentiation trajectory).
  • $g$: Net growth rate ($b - d$).
  • $b$: Birth rate.
  • Key Insight: The diffusion coefficient is state-dependent and explicitly coupled to the birth rate: $D(x,t) = \sigma^2 + 2\delta^2 b(x,t)$. This means stochastic variability is driven by cellular turnover (birth rate), not just net expansion.

Model Architecture & Training

• Neural Parameterization:
  • Velocity Field: Parameterized as the negative gradient of a scalar potential ($\mathbf{v} = -\nabla \phi$) using a Multi-Layer Perceptron (MLP) to ensure identifiability.
  • Birth Rate: Parameterized by a time-dependent MLP ($b_\theta$).
  • Death Rate: Assumed proportional to birth rate ($d = \alpha b$), where $\alpha$ is a fixed hyperparameter.
• Optimization:
  • Particle Simulation: Uses the Euler–Maruyama scheme to simulate weighted particles evolving through the learned SDE.
  • Loss Function: Minimizes the Sinkhorn divergence between predicted and observed snapshot distributions.
  • Regularization: Incorporates Wasserstein–Fisher–Rao (WFR) regularization to penalize excessive transport velocity and growth, replacing ad-hoc constraints like mass conservation.
• Identifiability Solution: By structurally coupling the diffusion coefficient to the birth rate, scDIVIDE resolves the ill-posed nature of jointly inferring velocity, growth, and diffusion from marginal data.

3. Key Contributions

1. Theoretical Novelty: First derivation showing that the diffusion coefficient in population dynamics is structurally coupled to the birth rate ($b$) rather than the net growth rate ($g$). This allows the model to distinguish between quiescent and high-turnover homeostatic states.
2. Methodological Advance: Development of scDIVIDE, a biologically informed machine learning framework that uses the "birth-diffusion coupling" as an inductive bias, eliminating the need for arbitrary assumptions like mass conservation.
3. Robustness: Demonstrates that this coupling significantly stabilizes training and improves predictive accuracy compared to decoupled models or existing methods.

4. Results

The framework was validated on both synthetic and real-world datasets:

• Synthetic Data (Three-Gene Toggle Switch):

  • Ground Truth: A system where gene B regulates birth rate, creating regions of high turnover and increased variance.
  • Performance: scDIVIDE accurately recovered the ground-truth dynamics and birth rates (Pearson $r = 0.985$).
  • Comparison: Outperformed four existing methods (TrajectoryNet, TIGON, VarRUOT, scDiffEq) in predicting future cell distributions (measured by holdout Wasserstein-1 distance).
  • Ablation: The full model (coupled + WFR) showed the lowest variance across random seeds, proving the synergy between the birth-diffusion coupling and regularization.

• Real Data (Mouse Hematopoietic Stem and Progenitor Cells - HSPCs):

  • Dataset: Time-series scRNA-seq of HSPCs differentiating in vitro (Days 0, 2, 4, 6).
  • Biological Insight: scDIVIDE inferred that immature HSPCs have lower birth rates (slower division kinetics) compared to intermediate states, consistent with prior biological observations.
  • Comparison with Alternatives:
    • Standard VarRUOT inferred high growth in terminal states (an artifact of quadratic cost functions).
    • Modified VarRUOT inferred high growth in immature states (contradicting biology).
    • scDIVIDE correctly identified the non-monotonic birth rate pattern (low in immature, high in intermediate, low in terminal).
  • Pathway Analysis: Gradient attribution analysis revealed that genes driving the birth rate were enriched in chemokine and cytokine signaling pathways, known regulators of hematopoietic proliferation and survival. Conventional cell-cycle scores failed to capture these specific division kinetics.

5. Significance

• Quantitative Dissection of Fate Decisions: scDIVIDE provides a new conceptual avenue to quantitatively dissect how division-associated noise shapes cell fate decisions in multicellular systems.
• Beyond Net Growth: It shifts the paradigm from modeling net population expansion to modeling turnover-driven stochasticity, offering a more accurate representation of homeostatic tissues and developmental processes.
• Generalizability: The framework is applicable to diverse single-cell datasets where understanding the interplay between proliferation, differentiation, and noise is critical (e.g., cancer, regeneration, development).
• Methodological Impact: It demonstrates how mechanistic biological constraints (derived from first principles) can be integrated into deep learning to solve core identifiability problems in dynamical systems inference.

In summary, scDIVIDE represents a significant step forward in single-cell trajectory inference by rigorously integrating the physics of cell division into the mathematical framework of stochastic differential equations, enabling the accurate reconstruction of cellular dynamics and division rates from snapshot data.