Learning unsupervised feature representations for single cell microscopy images with paired cell inpainting

Area of Science:

• Cellular Biology
• Bioimage Analysis
• Machine Learning

Background:

• Cellular microscopy images offer valuable biological insights.
• Extracting these insights requires defining specific image features or measurements.
• Automated feature design is crucial for efficient analysis of large microscopy datasets.

Purpose of the Study:

• To introduce a convolutional neural network (CNN) for automated feature design in fluorescence microscopy.
• To develop a self-supervised method for learning cell feature representations without labeled data.
• To demonstrate the utility of learned features for biological discovery.

Main Methods:

• A self-supervised convolutional neural network (CNN) was trained on a paired cell inpainting task.
• The CNN learned to predict fluorescence patterns in one cell based on another from the same image.
• This approach leverages image structure while controlling for cell morphology and imaging variations.

Main Results:

• The method successfully learned high-quality features describing protein expression patterns in single cells from yeast and human microscopy datasets.
• Learned features were effective for exploratory biological analysis, including proteome-wide clustering and quantification of protein localization and variability.
• The CNN demonstrated the ability to capture high-resolution cellular components.

Conclusions:

• The developed CNN method provides a generalizable approach for automated feature representation learning in multichannel microscopy images.
• Self-supervised learning using paired cell inpainting is effective for extracting meaningful biological information from unlabeled microscopy data.
• The learned features facilitate advanced biological analyses, enhancing our understanding of cellular processes.