Cell Patterns: The Structure Nobody Notices

Introduction: What Is a Tissue Actually Made Of?

Look at a tissue under a microscope and it can seem like a mess, thousands of cells squeezed together with no obvious plan. But that mess has structure. Cells don't sit randomly next to each other. They connect and arrange themselves in patterns that let a tissue grow, hold its shape, and function properly. When those patterns change, it's usually a sign that something else is changing too, like development progressing, or a disease taking hold.

The problem is that a tissue isn't one cell, it's thousands of them, all interacting at once. You can't understand the pattern by staring at a single cell. You need a way to describe the whole arrangement, and that is a much harder mathematical problem than it sounds.

Epithelial tissues are the classic example. These are the tightly packed sheets of cells that line and protect surfaces throughout the body. For years, the standard way to study them was simple: count how many neighbours each cell has. A cell touching five others gets labelled a "5-cell," one touching six gets labelled a "6-cell," and so on. If you plot the distribution of these numbers across the tissue, you get a rough fingerprint. It's a useful start, but it's also nowhere near enough.

Why Counting Neighbours Isn't Enough

Here's the issue: two tissues can have almost identical neighbour counts and still be organised completely differently. Picture two rooms, each furnished with the exact same set of chairs, tables, and shelves. Count the furniture in each room and the lists match perfectly. But walk into the rooms and they feel nothing alike, one is arranged in neat rows, the other in scattered clusters. In simple terms, the inventory is the same but the layout isn't. Cells work the same way. Knowing how many neighbours a cell has tells you nothing about which cells with which neighbour counts sit next to each other, or whether they cluster together or spread out evenly across the tissue.

Figure 1: Clustered vs. Uniform Tissue with Matching Histograms

So scientists started representing tissues as networks instead. Each cell becomes a point,

called a node, and a line connects it to every cell it touches. This is called a contact graph, and it's a genuine upgrade. Now you can ask questions about relationships between cells, not just isolated counts. Do certain cells cluster together? Are some regions more tightly connected than others?

But even a network has a blind spot. It only tells you who's touching whom. It says nothing about where cells actually sit in space. Two tissues can have the exact same contact graph while their cells are positioned completely differently, the same way two friend groups can have identical friendships while living in completely different cities.

This is where researchers picked up the problem, using a tool called Topological Data Analysis, or TDA.

Topology: Studying What Doesn't Change

Topology is the branch of maths that cares about structure over exact shape. A circle and a stretched out oval are different geometrically, different curvature, different measurements, but topologically they're the same, because both are just one unbroken loop. If you stretch it, twist it, or anything similar, as long as you don't tear it or glue new parts to it, topology says it hasn't changed.

TDA takes that idea and applies it to data. Instead of asking exactly where everything is, it asks what shapes and connections survive even as you change your perspective on the data. Applied to a tissue, that means looking past the exact geometry of individual cells and asking what structural patterns hold up no matter how you scale or rotate the image.

The specific technique used here is called persistent homology, and the idea behind it is genuinely neat. Imagine slowly growing connections in the tissue, first linking only the closest cells, then gradually allowing more distant ones to connect as you loosen the threshold. As this happens, structures appear: loops, clusters, etc. Some vanish almost immediately. Others stick around for a long time before disappearing. The ones that persist are treated as meaningful features of the tissue, the ones that flicker in and out are treated as noise.

The result of this process is recorded as a persistence barcode, a set of horizontal bars, where each bar represents one structure, and its length represents how long that structure survived as the connections grew. Short bars mean background noise. Long bars mean real organisation.

Figure 2: Tracking a Loop Through the Filtration

Building a Fingerprint Out of Barcodes

One barcode alone doesn't tell you much. The paper's real contribution is combining several different filtrations to capture different aspects of the tissue at once.

Two filtrations are built from the contact graph, called sub and sup, which track how cells with different neighbour counts cluster together. This catches structure that plain neighbour counting completely misses. A third is built from the positions of cell centres, using something called a Vietoris-Rips complex, which captures the physical, spatial layout of the tissue rather than just who's touching whom.

Together, these give a much richer picture. Instead of just showing how many neighbours a cell has, or who it’s connected to, It shows how those connections and positions combine into a larger structure 

There was a technical snag though. Images of tissues come in different sizes, with different numbers of cells and different scales. A bigger image shouldn't automatically look more complex just because it has more cells in it, that would be measuring the photograph, not the tissue. So the researchers built a careful normalisation process. They fixed the number of cells sampled from every image, and rescaled every barcode so it can be fairly compared to any other, regardless of the original image's size. They even proved, mathematically, that these adjusted measurements are stable, small changes to the input only produce small changes to the output, so the method isn't fragile or overly sensitive to noise.

Testing the Fingerprint

With this toolkit built, the researchers tested it on real epithelial tissue images from chickens and fruit flies, and compared them against centroidal Voronoi tessellations, or CVTs, mathematical models often used as stand ins for biological tissue because their cell shapes look very similar.

That's exactly the point they wanted to probe. Does looking similar mean being structurally similar? The answer, it turned out, was often no. Some CVT models reproduced a tissue's neighbour count distribution almost perfectly while completely failing to match its deeper spatial organisation. The positions of the cell centres told a different story than the surface level shapes did.

The method also succeeded where older tools struggled. Fruit fly wing tissue at two different developmental stages, just 24 hours apart, looks nearly identical using traditional measurements. TDA still found a distinguishing signature, differences in how "4-cells" were surrounded by 6 and 7 sided neighbours. Thai was a pattern invisible to simple counting but was picked up clearly through the topological analysis.

Adding Machine Learning

Finally, these topological fingerprints were fed into a Random Forest classifier, a machine learning model that makes decisions by combining many smaller decision rules, and asked it to sort tissue images into categories. When the topological variables were added alongside traditional measurements like mean neighbour count, classification accuracy improved. Instead of replacing the old approach, the topology was filling in the gaps the old approach couldn’t see. 

Figure 3: From Tissue Images to Classification

Conclusion

Counting a cell's neighbours tells you something. It just doesn't tell you enough. A tissue's identity isn't stored in any single number, but it's actually stored in how thousands of small local relationships add up into something bigger. Topology gives mathematicians a language for describing exactly that, not what a system looks like frozen in one instant, but what structure survives as you zoom out. 

That's a strange, but important idea to sit with. The same maths that separates a circle from a figure eight can also tell two seemingly identical strips of biological tissue apart, and it can do it more precisely than simply looking ever could.

Reference:

Atienza, N., Jimenez, M.-J., & Soriano-Trigueros, M. (2021). Stable topological summaries for analyzing the organization of cells in a packed tissue. Mathematics, 9(15), 1723. https://doi.org/10.3390/math9151723