Learning consistent subcellular landmarks to quantify changes in multiplexed protein maps.

Area of Science:

• Cellular and Molecular Biology
• Computational Biology
• Genomics and Proteomics

Background:

• Spatial context is crucial for understanding genome activity and cellular function.
• Highly multiplexed imaging offers a powerful approach to study these spatial relationships.
• Existing methods may struggle to consistently analyze complex molecular profiles across diverse cell types and conditions.

Purpose of the Study:

• To introduce CAMPA (Conditional Autoencoder for Multiplexed Pixel Analysis), a deep learning framework for analyzing multiplexed imaging data.
• To enable the identification and quantitative comparison of subcellular landmarks.
• To uncover how spatial organization and molecular composition influence cellular phenotypes and variability.

Main Methods:

• Development of a conditional variational autoencoder (CVAE) for learning molecular pixel profile representations.
• Clustering of learned representations to identify subcellular landmarks.
• Application of high-resolution multiplexed immunofluorescence to perturbed cellular systems.

Main Results:

• CAMPA successfully identifies consistent subcellular landmarks across heterogeneous cell populations.
• Quantitative analysis of landmark size, shape, molecular composition, and spatial organization is enabled.
• Changes in subcellular organization due to perturbations in RNA synthesis, processing, or cell size were revealed.
• Links between membraneless organelle composition and cell-to-cell variability in RNA synthesis were uncovered.

Conclusions:

• CAMPA provides interpretable cellular phenotypes from multiplexed imaging data.
• The framework accelerates the creation of multiscale atlases of biological organization.
• CAMPA aids in identifying the rules governing how cellular context shapes physiology and disease.