Generative machine learning unlocks the first proteome-wide image of human cells.

Area of Science:

• Cellular Biology
• Computational Biology
• Proteomics

Background:

• Cellular functions depend on protein spatial organization, but current imaging methods visualize only a fraction of the proteome.
• Thousands of proteins within a human cell necessitate advanced imaging or computational approaches for comprehensive study.

Purpose of the Study:

• To develop a deep generative model, ProtiCelli, capable of simulating microscopy images for a large number of human proteins.
• To create a large-scale dataset of virtual cells for advanced biological modeling and analysis.

Main Methods:

• ProtiCelli, a deep generative model, was trained on 1.23 million images from the Human Protein Atlas using three cellular landmark stains.
• The model simulates microscopy images for 12,800 human proteins.
• The generated dataset, Proteome2Cell, comprises 30.7 million images across 12 cell lines.

Main Results:

• ProtiCelli accurately reconstructs protein localization and preserves subcellular organization, outperforming existing methods.
• The model generalizes to unseen cell types and drug perturbations, infers drug effects from morphology, and predicts cell cycle stage.
• Generated images resolve compartment-specific functions and enable unsupervised segmentation of subcellular structures.

Conclusions:

• ProtiCelli enables spatial virtual cell modeling by computationally bridging the experimental scalability gap in spatial proteomics.
• The Proteome2Cell dataset facilitates the construction of hierarchical single-cell models and democratizes exploration of virtual cells.
• This work transforms spatial proteomics from protein cataloging to complete cellular system simulation.