cytoself

cytoself applies self-supervised deep learning to profile and cluster protein subcellular localizations from fluorescence microscopy images, enabling analysis of cellular architecture without preexisting annotations.

Key Features:

• Self-Supervised Learning: Employs a self-supervised training scheme that learns protein localization representations without manual labels or prior annotations.
• High-Resolution Protein Localization Atlas: Processes images of 1,311 endogenously labeled proteins from the OpenCell database to construct a multi-scale protein localization atlas capturing organelles and protein-complex signatures.
• Superior Clustering Performance: Demonstrates quantitatively validated superior clustering of proteins into organelles and protein complexes compared to previous self-supervised methods.
• Interpretability of Emergent Features: Provides interpretation of emergent features driving clustering by relating learned features to fluorescence image patterns.
• Component Analysis for Performance Optimization: Analyzes the contribution of individual model components to overall performance to identify which aspects drive success.

Scientific Applications:

• Cellular Architecture Mapping: Enables mapping of subcellular organization and identification of organelles and protein complexes from microscopy datasets.
• Protein Localization Profiling and Functional Inference: Supports profiling of protein localization patterns to study protein function and interactions within cellular compartments, informing hypotheses about roles in health and disease.

Methodology:

Uses self-supervised deep-learning training on fluorescence microscopy images from the OpenCell database to learn representations for profiling and clustering protein localizations, combined with quantitative clustering validation, interpretation of emergent image-based features, and component-level performance analysis.