Unveiling endothelial cell border heterogeneity: VE-cadherin adherens junction stratification by deep convolutional neural networks.

Background: Systemic diseases are often associated with endothelial cell (EC) dysfunction. A key function of ECs is to maintain the barrier between the blood and the interstitial space. The integrity of the endothelial cell barrier is maintained by VE-Cadherin homophilic interactions between adjacent cells. The morphology of these borders is highly dynamic and can be actively remodeled by numerous drivers in a (patho)physiologic context specific fashion.

Objectives: High-content screening of the impact of circulatory factors on the morphology of VE-Cadherin borders in endothelial monolayers in vitro will enable the assessment of the progression of systemic vascular disease. We therefore aimed to create an image analysis pipeline, capable of automatically analyzing images from large scale screenings, both capturing all VE-cadherin phenotypes present in a sample while preserving the higher-level 2D structure. Our pipeline is aimed at creating 1D tensor representations of the VE-cadherin adherence junction structure and negate the need for normalization.

Method: An image analysis pipeline, with at the center a convolution neural network was developed. The deep neural network was trained using examples of distinct VE-Cadherin morphologies from many experiments. The generalizability of the model was extensively tested in independent experiments, before further validation using ECs exposed ex vivo to plasma from patients with liver cirrhosis and proven vascular complications.

Results: Our workflow was able to detect and stratify many of the different VE-Cadherin morphologies present within the datasets and produced similar results within independent experiments, proving the generality of the model. Finally, by EC-cell border morphology profiling, our pipeline enabled the stratification of liver cirrhosis patients and associated patient-specific morphological cell border changes to responses elicited by known inflammatory factors.

Conclusion: We developed an image analysis pipeline, capable of intuitively and robustly stratifying all VE-Cadherin morphologies within a sample. Subsequent VE-Cadherin morphological profiles can be used to compare between stimuli, small molecule screenings, or assess disease progression.