StressNET: an adaptable deep-learning model for mechanical stress inference in tissues

Mechanical interactions between cells are fundamental to tissue morphogenesis during development and regeneration. Computational methods that infer intercellular stresses from microscopy images of cell shapes offer a non-invasive alternative to experimental perturbation techniques, yet all existing approaches rely on explicit physical models. Here we present StressNET, a Graph Neural Network (GNN) that infers intercellular mechanical stresses directly from tissue geometry, without assuming any underlying physical model. We generated synthetic datasets to train and benchmark StressNET, and demonstrate that its predictions achieve state-of-the-art correlation with experimental stress proxies in zebrafish neuromasts and Xenopus embryos. Analysis of the network's latent space reveals that StressNET learns global organizational principles of mechanical stress distribution, beyond local cell-cell interactions. StressNET is open-source and provides pre-trained models that can be fine-tuned on in vivo data, making it a broadly adaptable tool for studying tissue mechanics across biological systems.