Shape matters: Inferring the motility of confluent cells from static images

Cell motility in dense cell collectives is pivotal in various diseases like cancer metastasis and asthma. A central aspect in these phenomena is the heterogeneity in cell motility, but identifying the motility of individual cells is challenging. Previous work has established the importance of the average cell shape in predicting cell dynamics. Here, we aim to identify the importance of individual cell shape features, rather than collective features, to distinguish between high-motility (active) and low-motility (passive) cells in heterogeneous cell layers. Employing the Cellular Potts Model, we generate simulation snapshots and extract static features as inputs for a simple machine-learning model. Our results show that when the passive cells are non-motile, this machine-learning model can accurately predict whether a cell is passive or active using only single-cell shape features. Furthermore, we explore scenarios where passive cells also exhibit some degree of motility, albeit less than active cells. In such cases, our findings indicate that a neural network trained on shape features can accurately classify cell motility, particularly when the number of active cells is low, and the motility of active cells is significantly higher compared to passive cells. This work offers potential for physics-inspired predictions of single-cell properties with implications for inferring cell dynamics from static histological images.

💡 Research Summary

This paper tackles the challenging problem of inferring the motility of individual cells within dense, heterogeneous cell layers using only static microscopy images. The authors employ the Cellular Potts Model (CPM), a lattice‑based, cell‑resolved simulation framework, to generate two‑dimensional confluent monolayers composed of “active” (high‑motility) and “passive” (low‑motility) cells. Two distinct motility regimes are explored: (i) passive cells are completely non‑motile (κₚ = 0) while active cells experience a strong self‑propulsion force (κₐ = 1500), and (ii) both cell types are motile but active cells are considerably more motile (κₐ > κₚ > 0). By varying the number of active cells (Nₐ = 0–60) and the ratio γ = κₚ/κₐ, the authors create a rich dataset that mimics biological scenarios such as partial epithelial‑to‑mesenchymal transition (EMT).

From each CPM snapshot the authors extract 145 quantitative features, organized into four categories: (1) local shape (area, perimeter, major/minor axes, eccentricity, neighbor count, etc.), (2) non‑local shape (averages, minima, maxima of neighboring cells’ shape descriptors), (3) local structural metrics (bond‑order parameters ψₙ for n = 2…12, first and second moments of neighbor distances, distance standard deviation), and (4) non‑local structural metrics (neighbor‑averaged structural quantities). The feature set is deliberately interpretable, allowing a direct link between physical cell geometry and motility.

For classification, the authors use a simple multilayer perceptron (MLP) implemented in Scikit‑learn, with a single hidden layer containing as many neurons as input features. Weights are optimized with the ADAM algorithm. To avoid class imbalance, passive cells are randomly down‑sampled so that the training set contains equal numbers of active and passive cells. The dataset comprises 120 000 cell instances drawn from multiple independent snapshots; 80 % is used for training, 20 % for testing, and results are averaged over 20 independently trained networks.

Key findings:

1. In the κₚ = 0 regime, local shape features alone achieve >95 % classification accuracy, especially when the fraction of active cells is low (5–10 %). The strong self‑propulsion of active cells deforms their immediate neighborhood, producing distinctive elongation, eccentricity, and perimeter changes that are readily captured by static geometry.
2. When passive cells possess a modest motility (κₚ > 0), accuracy remains high only if the active‑to‑passive motility ratio is large (γ ≤ 0.2) and the active cell fraction is small. Under these conditions the MLP still reaches ~90 % accuracy. As γ increases or the active fraction grows, the geometric distinction blurs and performance drops sharply. Adding non‑local shape or structural features yields modest gains, confirming that most discriminative information resides in the cell’s own geometry.
3. A Gradient Boosting Decision Tree (GBDT) reproduces the MLP performance, whereas linear regression fails, underscoring the non‑linear nature of the shape‑motility relationship.

The study demonstrates that static, single‑cell morphological descriptors can serve as reliable proxies for dynamic motility phenotypes, provided the underlying physics (e.g., force balance, adhesion) is captured by the simulation. By coupling a physics‑based CPM with interpretable machine learning, the authors establish a proof‑of‑concept pipeline that could be transferred to real histopathology images. Potential applications include non‑invasive assessment of EMT‑driven cancer aggressiveness, quantification of immune cell activation in inflamed tissues, or monitoring of wound‑healing dynamics from fixed samples.

Future work should validate the approach on experimental data, extend the framework to three‑dimensional tissues, incorporate additional biophysical cues (e.g., substrate stiffness, chemotactic gradients), and explore temporal models that combine static morphology with limited dynamic information. Nonetheless, this paper provides a compelling demonstration that “shape matters”: the geometry frozen in a single image encodes enough information to infer how a cell moves within a crowded environment.