Optimizing lymphatic endothelial differentiation using metabolic pathway and machine learning

Introduction:: Induced pluripotent stem cell (iPSC)-derived endothelial cells (dEC) are crucial for developing vascularized tissues and modeling vascular diseases, but the existing differentiation protocols suffer from two major challenges. First, the protocols yield high percentage of undesired cell lines that share the mesodermal lineage such as fibroblasts, reducing the yield of dEC and necessitating removal of these cell lines to prevent fibroblast overgrowth. Second, there is a lack of efficient protocols that target specific subvariants of dECs, namely into lymphatic endothelial cells (LEC), which are useful for developing lymphatic organoids and promoting wound healing. iPSCs undergo changes to membrane potential as they differentiate, suggesting that a potential method to improve the yield of dEC may be to target membrane polarization of iPSC. Another method to improve the specificity of differentiation protocols into LEC may be to take advantage of its unique metabolic characteristics. LECs rely heavily on fatty acid oxidation for its metabolic needs and for generation of acetyl-CoA used for activation of VEGFR3, an LEC-specific marker with a positive feedback loop to a master lymphatic regulator gene, Prox1. Therefore, in this study, we explore the use of polarizing molecules and acetyl-CoA precursor molecule to improve both the yield of total dECs and specifically the LECs. Additionally, we explore the use of convolutional neural network coupled with machine-learning based segmentation algorithm to identify the desired dECs and LECs from imaging data without relying on immunofluorescence or flow cytometry data.


Materials and Methods:: Based on the existing differentiation protocol for dEC published by Harding and colleagues, we developed the following modified protocol (Fig 1A). We induce mesodermal identity through 2 days of treatment with 6μM of CHIR99021, a Wnt-pathway activator. Then we induce endothelial progenitor characteristics through 2 days of treatment with 50ng/mL of VEGF-A, FGF2, and BMP4 along with 10nM of ouabain, a small molecule that causes depolarization. Cells are passaged and replated to density of 50,000 cells/cm². Lastly, we treat the cells with 100ng/mL of VEGF-C and up to 50mM of sodium acetate (NaAc) for 4 days to induce LEC markers. To analyze the effects of our treatments, cells were passaged either after the endothelial progenitor induction stage or the LEC maturation stage. We also created a deep learning model using Keras library in Python to distinguish between endothelial cells and fibroblasts (Fig 1B). The model first processes the image into smaller segments centered around each cell identified using Cellpose library, which uses backbone U-net based neural network to accurately segment non-circular cells such as ECs. We then use each of the segments to train a network consisting of three convolutional layers and four dense layers, whose depth are optimized using Keras-tuner library. We pre-stained the ECs with fluorescent dye and re-plated stained ECs with unstained fibroblasts in 1:1 mixture at approximately 20,000 cells/cm² immediately before imaging. Each segment generated by the algorithm was labeled as either EC or fibroblast based on the presence of fluorescent signals within the cell boundary.


Results, Conclusions, and Discussions:: After 4 days, following the endothelial induction and treatment with ouabain, we performed flow cytometry for CD144, an endothelial marker, and Tra-1-60, a pluripotency marker. We observed a significant increase in the yield of CD144+ endothelial cells from approximately 50% to 70% when the iPSCs were treated with 10nM of ouabain during the endothelial progenitor induction stage (Fig 1C-D). All cells were Tra-1-60 negative and ouabain treatment during the mesodermal stage did not affect CD144+ yield, indicating that ouabain specifically affects cells in the mesodermal-endothelial state rather than drive iPSC in general towards differentiation.

After 8 days, we performed qPCR and flow cytometry for various LEC markers and CD31, a mature endothelial marker and found a significant increase in lymphatic markers PDPN and Prox1 when treated with 50mM NaAc (Fig 1E-G). We replicated the method across three iPSC cell lines and saw an increase in PDPN+/CD31+ LECs from approximately 5-10% to 40-60% with NaAc treatment (Fig 1H).

Using the optimal model, we confirmed that the accuracy of the model reaches approximately 80% (Fig 1I). The dataset was approximately 60% fibroblasts and 40% endothelial cells. We observed that the model begins to overfit by 200 epochs, which is when the validation loss reaches similar value to the training loss. We tested the optimized model on a testing image and generated the output image based on the predictions of the model (Fig 4J-K). We observe that the model accurately predicts most cells, but it misclassifies on cells that are closely packed due to difficulties in segmenting dense cell clusters.

Our differentiation protocol improves both the yield of CD144+ dECs and PDPN+/CD31+ mature LECs, which will be useful lymphatic and vascular tissue engineering research. We also developed a potential tool that can be used to distinguish dECs from the byproducts such as fibroblasts and fibroblast-like cells, which will help researchers quickly determine the efficacy of their differentiation protocols.


Acknowledgements (Optional): :

References (Optional): :