Self-supervised deep learning encodes high-resolution features of protein subcellular localization.

Area of Science:

• Cell biology
• Bioinformatics
• Artificial intelligence

Background:

• Understanding protein localization is key to deciphering cellular architecture and function.
• Existing methods often require extensive prior knowledge or annotations.

Purpose of the Study:

• To introduce Cytoself, a fully self-supervised deep learning approach for protein localization profiling and clustering.
• To generate a high-resolution protein localization atlas without relying on predefined categories or annotations.

Main Methods:

• Cytoself utilizes a self-supervised training scheme on protein images from the OpenCell database.
• The model processes images of 1,311 endogenously labeled proteins.
• Feature extraction and clustering performance were analyzed to understand model interpretability.

Main Results:

• Cytoself successfully generated a detailed protein localization atlas, reflecting cellular organization from broad categories to specific protein complexes.
• Quantitative validation demonstrated Cytoself's superior performance in clustering proteins into organelles and complexes compared to previous self-supervised methods.
• Analysis of emergent features provided insights into the model's internal workings and the basis of its clustering.

Conclusions:

• Cytoself offers a powerful, annotation-free tool for mapping protein localization with high resolution.
• The approach advances our understanding of cellular organization and protein complex distribution.
• Self-supervised deep learning holds significant promise for biological data analysis and discovery.