Collective Cell Motion in an Epithelial Sheet Can Be Quantitatively Described by a Stochastic Interacting Particle Model

When a wound opens in living tissue, thousands of cells have to decide, collectively, which way to move — and fast. How does a cell, that can only sense its immediate neighbors, end up part of a coherent, large-scale migration? That's not a vague biological question. Sepúlveda and colleagues at the interface of physics and cell biology sat down and answered it with a specific, testable model. The experimental setup is worth picturing clearly. The team worked with Madin-Darby Canine Kidney cells — MDCK cells — a standard workhorse for epithelial biology, arranged in a dense, confluent sheet. Instead of the classical wound assay where you scrape cells away with a pipette tip, they used microfabricated stencils: physical masks placed over portions of the culture surface. Pull the stencil off, and you expose a clean, precisely shaped strip of free surface. The cells don't know it's coming. Within minutes, the sheet begins to move. Over hours, thousands of cells migrate collectively into the gap, and the whole process is recorded as a time-resolved map of cell velocities, measured using Particle-Image Velocimetry — PIV — a technique borrowed from fluid mechanics that tracks motion the way weather radar tracks wind. What makes this collective motion puzzling is the scale of the coordination. Each cell interacts mechanically with only its immediate neighbors. Yet the whole sheet moves coherently across distances spanning hundreds of cells. Something local is producing something global. That's the question Sepúlveda and colleagues set out to answer quantitatively. Their model is deliberately minimal. Every cell is treated as a point particle. Two rules govern its motion. First, each cell is self-propelled but noisy — it moves stochastically, with a drive that fluctuates randomly over time. Second, each cell tends to nudge its velocity toward the velocities of its neighbors. In the mathematical formulation, the rate of change of a cell's velocity is a sum of four contributions: a noisy active drive — modeled as an Ornstein-Uhlenbeck process, meaning the noise has a finite memory rather than being purely instantaneous — plus a damping term representing friction against the substrate, plus an alignment term that pulls the cell's velocity toward each nearest neighbor's velocity, plus short-range mechanical forces that prevent cells from overlapping or drifting too far apart. That's it. Four terms. The neighbors themselves are defined geometrically: the surrounding space is divided into six sectors, and the closest cell within each sector, up to about 100 micrometers, counts as a neighbor. This kind of velocity-alignment rule is inspired by Vicsek-style flocking models — the same physics that describes how starlings murmurate. But the Sepúlveda formulation is explicitly stochastic and Langevin-like, without imposing a fixed preferred speed. It's a "wander and nudge each other" model, compact enough to simulate four thousand particles at once. The payoff comes when you put those simulations up against the PIV data at early times — about 30 minutes after stencil removal. The comparison is direct: roughly 4,000 simulated particles on a one-millimeter-square domain, matched to the experimental cell density, with the simulated velocity field compared against the measured one. The model nails the bulk statistics. The distribution of velocity components is close to Gaussian; the distribution of speed magnitudes follows a Lorentzian, and the simulation curves sit on top of the experimental points. Spatial velocity correlations — the degree to which two cells separated by a given distance tend to move in the same direction — match the experimental measurements including their decay over a characteristic length scale of around 500 micrometers. Temporal decorrelation happens on a timescale of roughly 5 hours in both the model and the experiment. Two rules, four thousand particles, and you get quantitative agreement with what living cells actually do. But then something happens at the advancing border that the bulk model alone cannot explain. The epithelium doesn't sweep forward as a smooth wave. It breaks up into finger-like protrusions — a digitated front — each one tipped by a single specialized cell that Sepúlveda and colleagues call a leader cell. These leaders are unmistakable in experiments: they are about three times larger than ordinary cells, often binucleated, non-dividing, and sporting an active lamellipodium — a flat, sheet-like protrusion — along their free-surface face. They move fast, with speeds peaked around 18 micrometers per hour. To capture this in the model, the team introduced a modified class of particles with two key differences from ordinary cells. First, leader cells interact with the free surface much more aggressively. In the model, unexplored free surface is represented by surface particles that repel ordinary cells, acting like a barrier to invasion. Leader cells hammer through that barrier far more effectively — their interaction with surface particles is both stronger and longer-range — so the repulsive shield is locally destroyed wherever a leader probes it, while ordinary cells remain effectively blocked. That asymmetry is what generates the fingers: leaders punch through, followers cannot. Second, leaders don't just run wild. They stay coordinated with the cells behind them. In the simulation, if a leader has four or more neighbors, it maintains its outward velocity. If it finds itself temporarily with fewer neighbors — meaning it has pulled ahead of its followers — it slows and adopts the mean velocity of those neighbors, up to a cap of 20 micrometers per hour. This feedback prevents leaders from detaching entirely from the sheet and is the ingredient that produces realistic, stable fingers rather than isolated escapees. The quantitative agreements here are striking. Simulated fingers show lower cell density than the bulk, decreasing continuously from base to tip — matching the experimental observation. The shapes of the digitated borders in simulation closely resemble those in the actual images. And the model does something more: it explains the regimes of border progression mechanistically. Each new leader entrains a portion of the border whose lateral extent is roughly the velocity correlation length, about 150 micrometers. At early times, when new leaders appear at a steady rate of about 0.15 leaders per hour per millimeter of border, the average border speed grows as time squared. Later, as that rate drops to around 0.05 per hour per millimeter, the border transitions to a roughly constant-speed regime. Both regimes were observed experimentally, and the model accounts for them without additional tuning. What this adds up to is a two-level explanation for collective epithelial migration. At the bulk level, stochastic self-propulsion plus neighbor-alignment is sufficient. That combination, with no more elaborate chemistry or mechanics, reproduces the full statistical character of how thousands of cells move together. At the border level, a small subpopulation — leaders — provides the spatial bias that sculpts the geometry of invasion. Sepúlveda and colleagues are explicit about where this sits among other modeling strategies. You can model epithelial sheets with detailed cell-shape descriptions — Potts models, vertex models — that allow you to incorporate biochemical details but are computationally expensive. You can use continuum descriptions that treat the tissue as a fluid or elastic medium. Particle-based models like this one occupy a middle ground: fewer parameters, cheaper to simulate, and easier to analyze mathematically, while still tracking individual cells and their interactions. Unlike some prior particle models, this one lacks a built-in preferred speed and does not spontaneously align without noise — it is a genuinely stochastic description, not a deterministic one with noise bolted on. The biological stakes are real. Collective migration of epithelial sheets underlies wound healing, embryonic development, and cancer invasion. Achieving a quantitative, minimal model of that migration — one specific enough to make predictions and fit data, not just reproduce qualitative trends — is a step toward understanding how to perturb it deliberately. Leader cells, in particular, are an obvious target: if the geometry of invasion depends on a small minority of cells with specific mechanical properties, then changing those properties changes everything downstream. The model gives you a handle on exactly what those properties need to be. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.