REGIS

Simulate, Don’t Interpolate

Recovering continuous spatial dynamics from discrete, unpaired snapshots

Simulated trajectories recovered from unpaired snapshots. Cycling MNIST: simulation of infinite cycling through the sequence of MNIST digits. Ising coarsening: a physical model of magnetism. Heart regeneration: simulation of cell activity after an in silico injury, learned from real spatial transcriptomics data.

Introduction

Many complex systems can only be understood through their change over time. Examples range from cell-fate decisions in biology[1] to chemical pattern formation[2] and grain growth in materials[3]. Modelling these systems requires understanding the processes that generate the observed states and govern their subsequent evolution. Yet complete trajectories are often difficult or impossible to observe. Destructive measurements, for example, provide detailed snapshots while preventing repeated observation of the same sample[4]. These observations provide a fragmented view of the system: they describe its marginal distributions at different times but do not directly reveal individual trajectories. Our aim is to learn a dynamical process from these unpaired snapshots, so that we can simulate how the system evolves and investigate its behaviour beyond the observed times or under local perturbations.

That is an ill-posed problem: distinct processes can produce the same distributions, so fitting the distributions alone cannot identify the dynamics that generated them[5]. Optimal transport[4], continuous flows[6] and stochastic bridges[7], flow matching[8] and stochastic interpolants[9], and recent multi-marginal and adversarially-learnt-interpolant variants[10][11][12][13] aim to reproduce the observed marginals, but the dynamics between them depend on modelling assumptions that the snapshots alone cannot determine. REGIS (REcovering Generative dynamics from Independent Snapshots) takes a different approach: it uses architectural constraints to narrow the space of rules that produce the unit system updates. These constraints come from our intuition about the modelled systems and, while being awkward to state about a probability path, they are often natural to state about an update rule. In this work, we focus on modelling spatial processes and therefore learn a rule that is local in space, Markovian and time-invariant, and which is applied iteratively to simulate the evolution of the system. We implement the rule as a neural cellular automaton[14][15] and train it adversarially[16] against the observed distributions, under the relativistic objective of R3GAN[17].

Below, we demonstrate that this approach can be applied to processes with very different dynamics: MNIST digits cycling through a sequence, a physical model of magnetism, and tissue regenerating after injury — learned from real spatial transcriptomics data. And because the rule is local and Markovian, the state can be edited part-way through the evolution, enabling in silico perturbations as experiments inside this framework without the need for any additional machinery.

t0 t1 t2 time marginal distribution matching at t2, and at every observed time ~ pobs ~ pgen zt zt+1 one generator update Generator One rule, applied everywhere and always in the same way zt+1= NCAθ(zt, noiset) Every cell reads features from its own 3×3 neighbourhood, passes them through one small shared network, and is updated by the result. The same NCAθacts at every cell and every step, so one rule has to account for the whole trajectory. θ θ zt θ θ zt+1 step the same θ at every cell and every step Discriminator marginal distribution matching Using the standard GAN adversarial game, the discriminator ensures that the snapshots from the generated trajectories match the observed ones at all available times. This procedure doesn't require the entire real trajectory to be observed, making it perfectly suitable for our setting. observed distribution generated distribution Training drives the generated distribution onto the observed one at each time. unobservable trajectory unpaired observation simulated trajectory
The structure of the data and the role of each component of REGIS. Time runs left to right. At each observed time we get samples from the marginal distribution associated with that time (blue); the trajectories these samples originate from are never observed in full (grey dashed). REGIS generates full trajectories (gold) by iteratively applying a local update rule, and the discriminator pushes the generated distribution onto the observed one at each observed time.

Stable long-horizon dynamics

A ring of ten positions like a lock dial, one per digit, with a stack of sample images of that digit at each position. A single circular arrow inside the ring runs clockwise from zero through nine and back to zero. An arrow leads out of the ring into a box labelled REGIS, and on to a panel showing a single digit deforming continuously into the next.
Cycle through the digits. REGIS learns a single local rule; iterating it produces a trajectory for an individual sample that guides it through the cycle. A full cycle from 0 through 9 and back to 0 requires 300 updates.

The handwritten digits in MNIST[18] have no natural time evolution, but we can define one: a cycle 0 → 1 → ⋯ → 9 → 0, learned as transitions from each digit to the next. The point of a cycle is that it never ends — once we have learned the rule, we can apply it sequentially and test the stability of the trajectories on ultra-long horizons. Starting from real test images we run up to 1000 consecutive transitions and measure the Fréchet distance[19] to real images in a frozen MNIST classifier’s feature space.

ModelParametersDigit transitions completed
15101001000
REGIS (ours)88 K 37.228.9–54.4 28.121.9–44.4 45.728.8–49.9 20.819.1–25.3 21.418.9–24.7
Non-local control 83 K 53.939.9–81.1 13155.7–270 208161–294 1,304†1,017–1,575 1,512†1,376–2,533
MMSFM 18.6 M 81.042.6–122 15986.7–215 46.626.9–66.6 262†160–1,272 34,928†5,498–64,357
MMtSBM 3.6 M 285195–429 303204–2,059 848184–6,176 1,596176–64,559 1,852†176–55,282
Image quality over repeated digit transitions. Fréchet distance in a frozen MNIST classifier’s feature space (lower is better), from 5000 generated and 5000 real samples at each horizon. Each cell is the median over seven seeds. † At least one seed diverged and is excluded.

All models in the comparison perform well in the beginning; however, the gap increases over time and at the horizon of hundreds of digit transitions, REGIS is the only model that retains its initial quality. Models also take distinct approaches to the transition between digits, as illustrated in the video below - while REGIS gradually reshapes the stroke, reusing the present structures, other methods merely interpolate between digits or noise-denoise them.

How the models move between digits. Four matched examples begin from the same real test zeros. REGIS (left), MMSFM (centre) and MMtSBM (right) follow the same continuous playback clock through ten digit transitions.

Run the cycle yourself

The entire model has just 88 K parameters, making it suitable for in-browser local inference. Sample a real test digit or draw your own, press play, and leave it running! The simulation can run forever.

If you decide to draw your own digit, say a 3, note that it isn't trivial that the system automatically knows that it should produce a 4 next. No external classifier is involved, and all iterations are strictly local; therefore the model should first gather the information in a distributed fashion to understand what has been drawn. Such self-classifying behaviour has been reported before[21], but it's still fascinating to see it in a dynamic system.

Cycling MNIST interactive widget 32×32 grid · 88 K parameters · 16 hidden channels not loaded

Loading fetches the learned rule (368 KB) and 80 real MNIST test digits (80 KB), so you can run the cycle locally.

0 full cycles
 
Start from
Run
Speed
Handwriting

The latent code is fixed during each run, but the updates are stochastic. Click above to change the code.

Data-efficient rule recovery and extrapolation

Cycling MNIST experiments show the model can hit a sequence of marginals, but this system lacks real underlying dynamics. To test whether REGIS can learn the process and not just copy the snapshots, we turn to the Ising model — a theoretical model of ferromagnetism that shows the behaviour of spins on a 2D lattice, evolving under Glauber dynamics[22]. We train REGIS on five marginal distributions, representing different timepoints in the coarsening process[23]: t ∈ {3, 10, 17, 300, 1000}.

Left, five unpaired Ising marginals at t = 3, 10, 17, 300 and 1000, each illustrated by a 128 by 128 crop, with a bar above marking the mean domain length growing across them. Below them, mean domain length against simulation time on log axes: the ground-truth growth law and the learned rule track one another through the observed window and on into the shaded region past the last observation, out to four times it. Right, one 512 by 512 lattice simulated by the same rule, four times the linear size it was trained on, with the 128 by 128 training crop drawn on it as a dashed square.
A model trained from unpaired data can extrapolate into larger images and beyond the training horizon. Left: 5 images, representing the state of the system at 5 training marginal timepoints t ∈ {3, 10, 17, 300, 1000}. The bar above each is the measured mean domain length at that time, drawn to the crop’s own scale. Below, mean domain length L against simulation time t on log axes, demonstrating that REGIS matched the ground-truth growth law. The shaded band is past the last observation, showing the extrapolation capability. Right: REGIS simulating the Ising coarsening process on a 5122 lattice. The simulation time in the left and right panels is synchronized.

Since the true dynamics are known, we can evaluate REGIS on a much deeper level compared to the MNIST experiment. Beyond just matching the growth law, REGIS exhibits the following properties:

Real world application recovers textbook biology

Many biological processes are inherently spatiotemporal: cell–cell interactions drive the formation and remodeling of tissue structure in development, regeneration, and diseases such as cancer. We therefore applied REGIS to a spatial atlas of adult zebrafish heart regeneration, which captures how the ventricle regrows after partial amputation [25]. Stereo-seq profiling of heart sections gives the spatial position of every annotated cell type in the uninjured heart and at seven timepoints after injury. The dataset is small — 4–6 hearts per timepoint, three sections per heart — and presents the typical obstacles of biological data: measurement is destructive, so paired observations are impossible, and heart shape varies considerably between individuals. The trained model simulates the otherwise unobservable trajectory of an individual heart's response to an in silico injury.

Left, the zebrafish heart atlas drawn as a loop: an uninjured state, a flat arrow marked injury leading to 6 hours post-amputation, and then a sling-shaped curve carrying 12 hours, 1, 3, 7, 14 and 28 days post-amputation back round to uninjured. Each timepoint hangs off the curve as a stack of real sections from different fish. Right, one section simulated by the learned rule, with a gold dot on the loop marking where the simulation has reached.
Learning heart regeneration from unpaired tissue sections. Left: real heart sections at the uninjured state and seven timepoints after amputation, arranged as a regeneration loop. The central diagram shows how the data are collected: hearts are injured, sampled at different times, sectioned and profiled. The immediate injured state is constructed by applying a manually annotated wound shape to an uninjured section; the dashed outline marks the removed tissue. Right: REGIS simulates regeneration in a single section, first after one virtual injury and then after another at a different location. Colour represents the dominant cell type. The simulation time in the left and right panels is synchronised.

Interactive heart regeneration simulator

As in the MNIST experiment, the model is small enough to run locally in your browser. Click on the blue region to injure the heart and watch it regenerate. The injury triggers a cascade in which border-zone cardiomyocytes (Bzm) activate, dedifferentiate, proliferate, and redifferentiate, accompanied by the appearance of stromal cells that help regulate regeneration. None of these steps were ever observed in the heart you are simulating: every change follows solely from the learned local, time-invariant update rule.

Heart regeneration loading

Loading the heart simulation…

Or draw your own cut. Drag on the blue ventricle, then press Play.

Uninjured
Ready for an injury
Cell activity and colour key

The plots track cell density as the heart regenerates. Each line uses the cell type’s colour in the key below.

Regeneration programme
Regulatory stroma

Colour shows the dominant cell type; opacity reflects tissue density. Border-zone myocardium (BZm) states are highlighted when they exceed their 99.7th-percentile thresholds.

The training data contain cuts to the ventricular apex. Cuts to the yellow bulbus arteriosus or red atrium fall outside those observations.

Does it reflect known biology?

Complete experimental trajectories cannot be observed, so we cannot compare them with the simulated ones directly. We can, however, test whether the model reproduces known biological relationships. Heart regeneration involves border-zone cardiomyocytes (Bzm), the heart muscle cells near the injury, together with macrophages, pro-regenerative fibroblasts, and epicardial cells. The cardiomyocytes pass through activation, dedifferentiation, proliferation, and redifferentiation — the sequence that defines the regeneration cascade [25]. If the learned dynamics reflect this sequence, suppressing one state should disrupt the states downstream of it, which gives a falsifiable prediction.

We test this prediction with in silico knockouts, setting the abundance of a selected cell population to zero after every simulation update. We then compare the cell composition at 28 days after injury with that of the unperturbed simulation and measure how the knockout changes the cardiomyocyte response. Valve cells serve as a negative control, since they are not involved in regeneration.

Knocking out regeneration-associated populations stops the system before the end of the cascade, leaving the simulated heart far from the healed state. Knocking out valve cells, in contrast, leaves simulated healing intact. The model was never trained on knockout experiments: these responses emerge from a rule learned only from unpaired snapshots.

Left, the regeneration loop with a dashed control trajectory just inside it that closes the same way, and a third that leaves sideways and never closes. Right, three rollouts side by side: unperturbed, with a negative control held at zero, and the border-zone dedifferentiation state held at zero, each over a plot of the five cascade states that fills in as the rollout reaches it.
Comparing regeneration with and without cell-state knockouts. Left: the regeneration loop with trajectories for three conditions. Right: simulations of regeneration starting from the same heart section and wound, either unperturbed or with a selected cell state set to zero after every update. Knockout of valve cells serves as a negative control, and its trajectory remains close to the unperturbed one. BZm (dediff.) is the dedifferentiating border-zone cardiomyocyte state; suppressing it reduces the activity of later states and disrupts the return to the uninjured state. The corresponding trajectory never heals the injury.

Conclusion

Learning a local, time-invariant rule lets us turn unpaired snapshots into a process we can simulate and perturb. In the examples above, this was enough to sustain long trajectories, extrapolate beyond the training data, and recover meaningful dependencies between cell states. As spatial technologies become widely available, we hope this approach will help us study processes whose individual trajectories cannot be observed.

Bonus: REGIS on non-regular grids

Replacing the 2D convolution in the perception layer with a graph convolution lets the same kind of rule run on arbitrary arrangements of cells. As a proof of concept, we train the cycling-MNIST task on digits placed on random packings of 200 discs. Neighbouring cells are connected by a Delaunay triangulation, and the Sobel filters are replaced by a least-squares estimate of each channel’s gradient over a cell’s neighbours (this becomes the vanilla NCA’s Sobel filter on a square lattice). Everything else in the generator is the same.

Cycling MNIST on different lattices. Trained on random disc packings only, the rule is run on held-out disc arrangements: hexagonal, square and jittered square lattices. Moreover, the simulation works even on a grid whose nodes move during inference. One trajectory per row, ten transitions.

Once trained, the rule is not tied to the grids it was trained on: at inference it carries over to regular lattices it has never seen, and even to grids whose nodes are moving.

References

  1. p53 Dynamics Control Cell Fate. Purvis, Karhohs, Mock, Batchelor, Loewer & Lahav. Science 336, 1440–1444, 2012. doi
  2. Transition from a Uniform State to Hexagonal and Striped Turing Patterns. Ouyang & Swinney. Nature 352, 610–612, 1991. doi
  3. Grain Growth in Four Dimensions: A Comparison between Simulation and Experiment. McKenna, Poulsen, Lauridsen, Ludwig & Voorhees. Acta Materialia 78, 125–134, 2014. doi
  4. Optimal-Transport Analysis of Single-Cell Gene Expression Identifies Developmental Trajectories in Reprogramming. Schiebinger, Shu, Tabaka et al. Cell 176, 928–943.e22, 2019. doi
  5. Fundamental Limits on Dynamic Inference from Single-Cell Snapshots. Weinreb, Wolock, Tusi, Socolovsky & Klein. PNAS 115, E2467–E2476, 2018. doi
  6. TrajectoryNet: A Dynamic Optimal Transport Network for Modeling Cellular Dynamics. Tong, Huang, Wolf, van Dijk & Krishnaswamy. ICML, 2020. pmlr
  7. Simulation-Free Schrödinger Bridges via Score and Flow Matching. Tong, Malkin, Fatras et al. AISTATS, 2024. pmlr
  8. Flow Matching for Generative Modeling. Lipman, Chen, Ben-Hamu, Nickel & Le. ICLR, 2023. openreview
  9. Stochastic Interpolants: A Unifying Framework for Flows and Diffusions. Albergo, Boffi & Vanden-Eijnden. JMLR 26(209), 1–80, 2025. jmlr
  10. Modeling Complex System Dynamics with Flow Matching across Time and Conditions. Rohbeck, De Brouwer, Bunne et al. ICLR, 2025. iclr
  11. Multi-Marginal Stochastic Flow Matching for High-Dimensional Snapshot Data at Irregular Time Points. Lee, Moradijamei & Shakeri. 2025. arXiv:2508.04351
  12. Multi-Marginal Temporal Schrödinger Bridge Matching from Unpaired Data. Gravier, Boyer & Genovesio. ICML, 2026. arXiv:2510.01894
  13. Multi-Marginal Flow Matching with Adversarially Learnt Interpolants. Kviman, Tamogashev, Branchini, Elvira, Lagergren & Whitammer. ICLR, 2026. arXiv:2510.01159
  14. Growing Neural Cellular Automata. Mordvintsev, Randazzo, Niklasson & Levin. Distill, 2020. distill.pub
  15. Learning Spatio-Temporal Patterns with Neural Cellular Automata. Richardson, Antal, Blythe & Schumacher. PLOS Comput. Biol. 20, e1011589, 2024. doi
  16. Generative Adversarial Nets. Goodfellow, Pouget-Abadie, Mirza et al. NeurIPS, 2014. neurips
  17. The GAN is Dead; Long Live the GAN! A Modern GAN Baseline. Huang, Gokaslan, Kuleshov & Tompkin. NeurIPS, 2024. neurips
  18. Gradient-Based Learning Applied to Document Recognition. LeCun, Bottou, Bengio & Haffner. Proc. IEEE 86, 2278–2324, 1998. doi
  19. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. Heusel, Ramsauer, Unterthiner, Nessler & Hochreiter. NeurIPS, 2017. neurips
  20. Multi-Scale Context Aggregation by Dilated Convolutions. Yu & Koltun. ICLR, 2016. arXiv:1511.07122
  21. Self-classifying MNIST Digits. Randazzo, Mordvintsev, Niklasson, Levin & Greydanus. Distill, 2020. distill.pub
  22. Time-Dependent Statistics of the Ising Model. Glauber. J. Math. Phys. 4, 294–307, 1963. doi
  23. Theory of Phase-Ordering Kinetics. Bray. Advances in Physics 43, 357–459, 1994. doi
  24. Denoising Diffusion Probabilistic Models. Ho, Jain & Abbeel. NeurIPS, 2020. neurips
  25. An Organ-Wide Spatiotemporal Transcriptomic and Cellular Atlas of the Regenerating Zebrafish Heart. Li, Lu, Guo et al. Nature Communications 16, 3716, 2025. doi