Unbiased single-cell morphology with self-supervised vision transformers.

Area of Science:

• Computational Biology
• Bioimage Analysis
• Machine Learning

Background:

• Accurate quantification of cellular morphology is crucial for advancing single-cell biological studies.
• Existing computer vision methods for cell morphology analysis often require manual annotations or extensive supervision.
• Developing scalable and automated methods for morphological analysis remains an active research area.

Approach:

• We investigated the utility of DINO (self-supervised vision-transformer), a deep learning model, for learning cellular morphology representations.
• DINO was evaluated on diverse, publicly available imaging datasets without any manual annotations.
• The algorithm's ability to capture morphological features across multiple scales and identify sources of variation was assessed.

Key Points:

• DINO demonstrates a strong capacity for learning rich cellular morphology features in a self-supervised manner.
• The algorithm successfully encodes biologically and technically relevant information at subcellular, single-cell, and multi-cellular levels.
• DINO effectively distinguishes between biological and technical factors influencing image data.

Conclusions:

• Self-supervised learning with DINO offers a powerful, annotation-free approach to morphological analysis in biological imaging.
• DINO facilitates the study of complex biological variations, including single-cell heterogeneity and sample-level relationships.
• This method serves as a valuable tool for accelerating image-based biological discovery and understanding cellular phenotypes.