Graph neural network that infers intercellular mechanical tension from segmented epithelial geometry, pretrained on synthetic vertex-model tissues.
No providers recorded yet. Browse all providers
Cells in an epithelium pull on one another through their shared junctions, and those tensions shape how tissues fold, stretch and regenerate. Measuring them directly means laser ablation, magnetic droplets, optical traps or tension-sensitive probes, all of which perturb the tissue and yield few measurements. Computational stress inference instead reads cell shapes from a microscopy image and solves for the tensions that produced them, but every existing geometric method does so by writing down an explicit mechanical model and inverting a system of force-balance equations.
StressNET drops the explicit physics. It treats a segmented tissue as a graph in which each cell-cell interface is a node, and uses a graph neural network to regress the relative tension on every interface directly from geometry. Twelve layers of message passing and a global attention-pooling step give each junction tissue-scale context that junction-by-junction force balance ignores. It is trained entirely on simulated tissues, then applied to real embryos and larvae.
StressNET was developed by Nicolás Aldecoa Rodrigo and Augusto Borges with the Chara group at the University of Nottingham and the Universidad Argentina de la Empresa, with the López-Schier (zebrafish) and Sedzinski (Xenopus) labs contributing the in vivo data. It was posted to bioRxiv in September 2026 and has not yet been peer reviewed. It builds on the same group's physics-based ForSys method, which serves as its main baseline.
The input is a skeletonised segmentation converted into a polygonal mesh with the ForSys library. The network has 727,137 trainable parameters in four parts: a two-layer edge-feature extractor (512 to 128 dimensions); a 12-layer backbone of custom StressConv layers, extensions of the CrystalConv gated graph convolution, arranged in four blocks of widths 32, 48, 64 and 96 with residual add-and-norm; a GlobalAttnSumPool layer; and a three-layer regression head. Outputs are normalised to a mean of one; the implementation uses TensorFlow and Spektral.
Training data were 14 stress-pattern families of 100 tissues each, simulated with a 2D vertex model in Surface Evolver. In silico it recovered ground-truth tensions, tolerated vertex jitter mimicking segmentation error, and generalised to held-out pattern families. The in vivo tests separate zero-shot and fine-tuned use. On 15 laser-ablation experiments in zebrafish lateral-line neuromasts, the five synthetic-only models, with no fine-tuning, correlated with junction recoil velocity at a mean Pearson r of 0.813 ± 0.057, against 0.548 for ForSys. On 126 Xenopus mucociliary epithelium images using a myosin II sensor as the tension proxy, the synthetic-only models reached a median Pearson r of 0.39 versus 0.18 for ForSys. Fine-tuning on 88 Xenopus images raised the median correlation on a 38-image held-out set to 0.58 (0.59 with rotation-augmented inference); as few as 11 images lifted the mean from 0.29 to 0.44, and gains scaled logarithmically with sample count.
StressNET is aimed at developmental and cell biologists who image epithelia with a membrane marker and want a tension map without perturbing the tissue. Given a segmented image, it assigns a relative tension to every junction, which can be compared against myosin localisation or tracked across conditions. The stressnet Python package loads the pretrained weights, runs inference on Surface Evolver outputs or skeletonised microscopy images, plots stress maps, and fine-tunes on a user's own labeled data.
The authors present StressNET as the first deep-learning approach to geometric stress inference, a field otherwise built on inverting force-balance equations. Its main contribution is showing that a model trained only on simulated tissues transfers to two unrelated vertebrate epithelia and beats a physics-based inverse method there. Linear probes on its pooled embeddings recover stress-pattern family and tissue size, indicating it learns tissue-scale organisation. Limits are clear: it works on 2D apical geometry only, its outputs are relative rather than absolute tensions, validation rests on indirect proxies (recoil velocity and myosin intensity) across a modest number of samples, and 3D and time-resolved extensions remain future work.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.