Convolutional Neural Networks Can Predict Retinal Differentiation in Retinal Organoids

Area of Science:

• Biotechnology
• Stem Cell Biology
• Computational Biology

Background:

• Stem cell-derived organoids are crucial for studying tissue development in vitro.
• Assessing retinal differentiation typically requires invasive methods like reporter gene expression.
• A non-invasive, robust method is needed to predict differentiation outcomes.

Purpose of the Study:

• To develop a deep learning-based computer algorithm for predicting retinal differentiation in organoids.
• To utilize bright-field imaging for non-invasive assessment, avoiding chemical probes or reporter genes.
• To demonstrate the capability of convolutional neural networks (CNNs) in analyzing tissue development.

Main Methods:

• Trained CNNs (ResNet50v2, VGG19, Xception, DenseNet121) using transfer learning on labeled bright-field images of organoids.
• Classified organoids into 'retina' and 'non-retina' categories based on fluorescent reporter gene expression.
• Compared CNN performance against human expert classification.

Main Results:

• The best-performing CNN model (ResNet50v2) achieved an area under the curve of 0.91.
• The CNN algorithm correctly predicted organoid fate in 84% of cases, surpassing human experts (67%).
• Retinal differentiation was accurately predicted before reporter gene expression onset.

Conclusions:

• Deep learning algorithms can effectively recognize and predict retinal differentiation in stem cell-derived organoids using bright-field imaging.
• This approach offers a non-invasive, accurate, and efficient method for assessing organoid development.
• This study represents the first successful application of CNNs for classifying stem cell-derived tissues in vitro.