Sun, H., Kahnert, K., Hansen, J. N., Leineweber, W. D., Li, M., Feng, W., Ballllosera Navarro, F., Axelsson, U. et al.
This paper presents the first model to simulate the spatial distribution of the entire proteome at the single-cell level using generative machine learning.
Current imaging technologies can only visualize tens of proteins simultaneously within a single cell, creating a fundamental scalability gap in understanding the spatial organization of the thousands of proteins that govern cellular functions.
The authors developed ProtiCelli, a deep generative model trained on 1.23 million images from the Human Protein Atlas. It takes images of just three cellular landmark stains as input and simulates microscopy images for 12,800 human proteins.
ProtiCelli outperforms existing methods in reconstruction accuracy and texture fidelity, generalizing to unseen cell types and drug perturbations. It enabled the creation of Proteome2Cell, a dataset of 30.7 million simulated images, and demonstrated capabilities like inferring drug-induced changes from cell morphology and predicting cell cycle stages without dedicated markers.