3D: Orientation order in epithelial monolayers, towards a paper
Elective project, round 2 (P3). Report due Sun 23 May, defense Tue 25 May
The question
Over the last decade, many papers have reported nematic order in layers of epithelial cells: regions where the elongated cells align, with topological defects between them, analysed with the theory of active liquid crystals. Most of these claims rest on still images, and few compare the measured order with what a random layer would give. In 2025, two students in this course, Giulio Simonetti and Tobit Roth, built a full analysis pipeline for our quantitative phase movies of MDCK monolayers, from scratch. We have started a manuscript from it, with them as co-authors. What is missing is the part that decides what can be claimed:
Is there nematic order in MDCK monolayers that is stronger, larger or longer-lived than chance? At which densities, length scales and time scales?
This is a research project with an open answer. It continues 2C Active cell layers: if you did 2C you already have the tools; if not, the first week brings you up to speed. If the results are good, they go into the paper, and students who contribute substantially are offered co-authorship too.
What exists
- Time-lapse quantitative phase (Holomonitor) movies of MDCK monolayers from five experiments in 2024: freely expanding layers and layers confined on circular islands, 110–337 frames each, at a range of densities. The pixel values are cell heights.
- A Python package by Simonetti and Roth (segmentation, cell shapes and orientations, coarse-grained nematic fields, ±½ defect detection), extended in 2026 with pixel-resolution fields and frame-to-frame defect matching. Four notebooks take you from raw images to defect statistics.
- A manuscript draft, Persistence of elongation and orientation in MDCK monolayers, with the methods written and the results section planned in four parts: aspect-ratio distributions, the lifetime of cell orientation, nematic order against coarse-graining scale, and defect dynamics.
Data, code and notebooks are in one repository on UiO’s GitHub (github.uio.no/dagkd/orient_mdck), with five movies ready to analyse. Its README explains how to start. You get access, and access to the manuscript, at the kick-off.
Milestones
| When | What you show | |
|---|---|---|
| M1 | Tue 4 May (W18) | The pipeline running on at least two data sets, and one result from 2025 reproduced; your part of the paper chosen (below) |
| M2 | Tue 11 May (W19) | Validated core figure: your result compared with a null model, and the method tested on simulated data with known order |
| M3 | Sun 23 May | Report and LLM log on Canvas. The report is written as a draft results section of the paper, with its figures |
| M4 | Tue 25 May (W21) | Oral defense |
Choose one part of the paper
Each part is a project in itself. With two or three students, each takes a different part.
A. Cell shape: is the aspect-ratio distribution universal? Sadhukhan and Nandi (eLife 2022) predict a one-parameter distribution of cell aspect ratios for confluent layers. Measure the distributions for all data sets, test the predicted data collapse and the relation between the spread and the mean, and find how the parameter depends on cell density. Which physical quantity does it correspond to?
B. How long do cells keep their shape and orientation? Track individual cells from frame to frame and measure the autocorrelation times of their area, height, aspect ratio and orientation. Are shapes and orientations short-lived compared with other processes in the layer, such as the 2–4 hour height pulsations we see? If an orientation lasts only minutes, can it matter biologically?
C. Nematic order against length and time scale. Measure the coarse-grained order parameter \(S(R)\) and the orientation correlation function against distance and time lag. Compare each with a null model at every scale. Over what range, if any, is there order beyond chance, and how does it depend on density?
D. Are the defects real? Measure the density, lifetime and motion of ±½ defects against the coarse-graining radius. Do the same for orientation fields from the null model. Real defects should live longer, and move differently, than the defects that noise alone creates.
Validation: what makes a result publishable
- A null model for every statistic. Keep the positions and shapes of the cells but shuffle their orientations (within a frame, and between frames). Also compute the finite-sample baseline: with \(n\) cells in a window, random orientations give \(S\approx\sqrt{\pi/(4n)}\). A result counts only where it lies clearly outside the null distribution. In a related project on gastruloids, this test removed most of the long-range order that had first been reported.
- A test on known data. Apply exactly the same pipeline to simulated orientation fields with a known correlation length (or to Vicsek simulations at high and low noise, as in 2C). The pipeline must find the order you put in, and none where there is none.
- Robustness to segmentation. Rerun your main result with different segmentation parameters, and on a manually checked region. A result that changes when the segmentation changes is a result about the segmentation.
Working with an LLM on this project
This is research code that others will reuse, so it must be readable, tested and documented. An LLM can help you extend the package, write tests and draft text for the paper. It cannot judge whether a measured order is real. The null models are your job. Keep a log of the parameter choices that change the result, because reviewers will ask.
The oral defense
About 15 min per student. Typical questions:
- What is nematic order, and how is it different from polar order? Why use \(e^{2i\theta}\)?
- What is a ±½ topological defect? Why can noise create them, and how do you tell them apart from real ones?
- Explain your null model. Why is it the right one for your question?
- How does your result depend on the coarse-graining radius, and why?
- What would you need to claim that MDCK monolayers are an active nematic?
- What did the LLM get wrong?