Deep Learning Powered Identification of Differentiated Early Mesoderm Cells from Pluripotent Stem Cells

Area of Science:

• Stem cell biology
• Developmental biology
• Computational biology

Background:

• Pluripotent stem cells differentiate into three germ layers: ectoderm, endoderm, and mesoderm.
• Early identification and characterization of germ layers during differentiation are challenging, limiting high-throughput screening.
• Accurate lineage specification is crucial for regenerative medicine.

Purpose of the Study:

• To develop deep learning (DL) methodologies for predicting and classifying early mesoderm cells.
• To utilize cellular and nuclear morphologies for classification of mesoderm cells differentiated from embryoid bodies (EBs).
• To establish a rapid and high-throughput method for assessing lineage specification efficiency.

Main Methods:

• Utilized a transgenic murine embryonic stem cell (mESC) line (OGTR1) with mesodermal gene (Brachyury (T): DsRed) validation.
• Trained a convolutional neural network (CNN) model (InceptionV3) on phase-contrast and nuclear images for cell classification.
• Employed an Attention U-Net CNN for image segmentation of phase-contrast and nuclear images.

Main Results:

• The InceptionV3 CNN achieved 97% accuracy for phase images and 90% for nuclear images in classifying mesodermal vs. non-mesodermal cells.
• Attention U-Net CNN achieved mean intersection over union scores of 61% for phase-contrast images and 69% for nuclear images.
• Validated mesodermal gene upregulation (Brachyury (T): DsRed) in differentiated cells.

Conclusions:

• Deep learning models can accurately predict mesoderm cell differentiation based on morphology.
• Integration of cell culture, imaging, and DL accelerates lineage specification identification.
• This approach holds significant potential for advancing regenerative medicine through efficient screening.