SubCell: Proteome-aware vision foundation models for microscopy capture single-cell biology.

Area of Science:

• Cell biology
• Biophysics
• Computational biology

Background:

• Cell morphology and protein organization are crucial for understanding cellular function.
• Fluorescence microscopy generates large datasets for studying these features.
• Machine learning (ML) offers powerful tools for interpreting microscopy images.

Purpose of the Study:

• Introduce SubCell, a self-supervised deep learning model suite for fluorescence microscopy.
• Accurately capture cellular morphology, protein localization, organization, and function.
• Develop a proteome-wide hierarchical map of protein organization from image data.

Main Methods:

• Trained deep learning models on the Human Protein Atlas proteome-wide image collection.
• Utilized a novel proteome-aware learning objective for model training.
• Integrated SubCell with protein sequence models for multimodal representation.

Main Results:

• SubCell accurately captures cellular features beyond human perception.
• Outperformed state-of-the-art methods in single-cell biology tasks.
• Constructed the first proteome-wide hierarchical map of protein organization from image data.
• Achieved generalization to diverse fluorescence microscopy datasets without fine-tuning.
• Enabled comprehensive gene function capture through multimodal protein representation.

Conclusions:

• SubCell provides deep, image-driven representations of cellular architecture.
• The developed cell map defines subsystems, reveals protein functions, and distinguishes cellular behaviors.
• SubCell's multimodal approach enhances gene function understanding.
• The models are applicable across diverse biological contexts and datasets.