A deep learning approach to predict differentiation outcomes in hypothalamic-pituitary organoids

Area of Science:

• Stem cell biology
• Biotechnology
• Artificial intelligence in medicine

Background:

• Three-dimensional culture of human pluripotent stem cells generates pituitary organoids but faces challenges in maintaining consistent differentiation efficiency due to inherent cellular heterogeneity.
• The empirical nature of organoid culture processes necessitates objective methods for monitoring differentiation progress.

Purpose of the Study:

• To develop and validate a deep learning model capable of predicting the appropriate differentiation of pituitary organoids using image analysis.
• To assess the model's predictive accuracy compared to expert human observers.

Main Methods:

• Utilized three-dimensional culture systems of human pluripotent stem cells to generate pituitary organoids.
• Employed deep learning models, specifically EfficientNetV2-S and Vision Transformer, trained on bright-field images of organoids.
• Incorporated VENUS-coupled RAX expression as a marker for differentiation status and utilized ensemble learning for model improvement.

Main Results:

• The developed deep learning models achieved 70% accuracy in classifying bright-field images of organoids into three differentiation categories, surpassing expert predictions.
• Ensemble learning enabled the model to predict RAX expression even in cells lacking the RAX::VENUS reporter, indicating broader applicability.

Conclusions:

• Deep learning offers a powerful, accurate, and objective method for monitoring pituitary organoid differentiation.
• The validated model shows potential for clinical applications, including guiding cell transplantation strategies.