Predictive machine learning models for tissue mechanics and function

Mechanochemical feedback and force responses drive complex tissue behaviors. One of the main obstacles in uncovering the mechanisms of such complex tissue behaviors is the lack of systematic computational tools that provide insight into precise mechanotransduction events. Here, we develop “biologically-informed neural networks” (BINNs) to model and predict tissue-scale mechanotransduction phenomena through carefully curated domain-informed datasets and biologically relevant features. Using mechanochemical information contained in high-resolution images of the actomyosin network and VE-Cadherin-marked cell junctions across endothelial tissues, we trained U-Net convolutional neural network (CNN) models for predicting junction outlines as well as their morphologies/types, including linear and discontinuous junction classes, which are well-described readouts of healthy and atherosclerotic tissues respectively. We further trained this model with fluorescent microscopy datasets of junction morphologies coupled to spatial tissue permeability in endothelial tissues with a wide range of morphologies. This integrative BINN pipeline provides accurate predictions of the functional state of endothelial tissues. Overall, our ML approach provides a quantitative framework to define tissue-scale (dys)function and will be broadly generalizable to several mechanobiology questions.