What is it about?
Many tissue simulations model biology one cell at a time. A popular method, the Cellular Potts Model, can reproduce realistic cell movements and tissue patterning, but it can be slow—especially when you need to simulate long time periods or run many repeats. In this paper we train a deep-learning “surrogate” (a fast stand-in) that learns to predict how a Cellular Potts simulation evolves. We treat the task as pattern-evolution prediction: given the current simulated tissue and chemical signals, the model predicts what they will look like 100 simulation steps later. This enables much faster runs while preserving key short-term pattern-formation behaviors in a vasculogenesis (early blood-vessel formation) test case.
Featured Image
Photo by FlyD on Unsplash
Why is it important?
Faster simulations matter because researchers rarely rely on a single run. To trust a model, you often need to repeat simulations many times, try different settings, and see how stable the results are. If each run is slow, it becomes impractical to simulate large tissues, explore longer time periods, or do systematic “what-if” testing. Our approach provides a large speedup while still preserving the short-term pattern changes that are important in our test case: a lab-motivated model of early blood-vessel formation. In particular, the accelerated simulations can reproduce features such as the growth and joining of vessel-like structures and the filling-in of small gaps in the network. Overall, this helps make proven, mechanistic tissue models—models that represent tissues using explicit rules for how cells move, stick, and signal—more usable at the scales and time horizons needed for broader studies.
Perspectives
I’ve spent much of my career building mechanistic, multiscale models and developing CompuCell3D, a platform for cell-based tissue simulations that we also use to generate the simulations in this paper. One persistent limitation of detailed mechanistic models has been speed: even when the biology is represented in a transparent, cause-and-effect way, it can be hard to run the model enough times—or far enough forward in time—to use it as a practical predictive tool. This paper is fundamentally about the forward problem: given a current tissue state (and its chemical signals), can we reliably predict how the simulation will evolve over time—quickly enough to be useful? That capability matters because you can’t responsibly talk about designing interventions until you can run the forward model efficiently and repeatedly. Faster forward prediction is the foundation for exploring “what-if” scenarios, comparing alternative mechanistic assumptions, and eventually optimizing candidate interventions. What excites me most is the longer-term path this opens up. If we can make forward simulation fast while preserving the essential tissue-scale behaviors, we can start to imagine these models being embedded in workflows for personalized forecasting in complex biological systems—where you update the model to reflect an individual’s state and then predict how the system is likely to develop over time. Even if that future requires additional work (validation, calibration, uncertainty quantification), accelerating the forward dynamics is a key enabling step toward making mechanistic virtual tissues genuinely actionable.
James Glazier
Indiana University Bloomington
Read the Original
This page is a summary of: Surrogate modeling of Cellular-Potts agent-based models as a segmentation task using the U-Net neural network architecture, PLoS Computational Biology, November 2025, PLOS,
DOI: 10.1371/journal.pcbi.1013626.
You can read the full text:
Resources
Contributors
The following have contributed to this page
