A Machine Learning Framework to Predict Subcellular Morphology of Endothelial Cells for Digital Twin Generation

Abstract

Gaining insight into different cell behaviors is key to better understanding different pathologies. These behaviors may be explained in part through close observation of 3D cell morphology. Therefore, the objective of this research was to develop a machine learning (ML) framework that can predict 3D subcellular morphological variation of endothelial cells (ECs) to generate digital twins. ECs were cultured and their membrane, nucleus, and focal adhesion (FA) sites were stained and imaged with confocal microscopy. The multicellular confocal z stacks were segmented resulting in a total of 60 single-cell stacks. Fifty randomly picked cells were augmented 20-fold to train the ML framework, and the remaining 10 were used for an independent test of prediction accuracy. The ML framework was based on an open-source conditional generative adversarial network (cGAN), which was expanded to make 3D predictions using membrane only as input to predict nucleus and FA morphology. After training the framework, the results on the independent test showed an average prediction accuracy of ~87% for nucleus and ~70% for FA sites. The predictions were used to build a digital twin of each EC and compared to their respective ground truth, showing an average ~79% global accuracy and ~84% accuracy in FA-Nucleus distribution. The results presented show the effectiveness of the developed ML framework to generate digital twins of ECs using limited amount of data. These digital twins can be used to couple EC morphology with different behaviors. The ML framework can be potentially expanded to predict morphology of other subcellular structures as well as to study other types of cells.