A novel machine learning based approach for iPS progenitor cell identification.

Area of Science:

• Stem cell biology
• Computational biology
• Biotechnology

Background:

• Identifying induced pluripotent stem (iPS) progenitor cells is crucial for understanding reprogramming mechanisms.
• Early progenitor cells are rare (<5%) and lack known biomarkers, making experimental identification challenging.
• Current methods can only identify iPS cells around day 6, missing early progenitor dynamics.

Purpose of the Study:

• To develop a novel computational approach for identifying iPS progenitor cells during the early stages of reprogramming.
• To overcome the limitations of experimental identification due to low cell ratios and lack of biomarkers.
• To analyze cell morphology and motion patterns for distinguishing progenitor cells from normal fibroblasts.

Main Methods:

• Utilized live-cell imaging to record the reprogramming process of murine embryonic fibroblasts (MEFs) after viral infection (Oct4, Sox2, Klf4).
• Calculated 11 morphological and motion features from time-lapse microscopy images for cells 3-5 days post-infection.
• Developed a machine learning model (XGBoost) using selected features and optimal time windows for prediction, handling missing data.

Main Results:

• The XGBoost model achieved a minimum precision above 52% in identifying iPS progenitor cells within 3-5 days post-infection.
• Feature selection identified six key morphological and motion features crucial for accurate prediction.
• The model demonstrated robustness, applicable to datasets with missing values or frames.

Conclusions:

• The developed computational method effectively identifies iPS progenitor cells using machine learning and image analysis.
• Distinct morphology and motion patterns differentiate iPS progenitor cells from normal MEFs, enabling computational identification.
• This approach provides valuable insights into early reprogramming dynamics and the origin of iPS cells.