Spatial simulations of cell processes require realistic cell geometries — accurate representations of where organelles are, how they’re shaped, and how they vary from cell to cell. Building those geometries by hand doesn’t scale. CellOrganizer was designed to learn them directly from microscope images.
This chapter describes the full workflow: from image preparation through model training, quality assessment, geometry sampling, and integration with biochemical simulation frameworks.
CellOrganizer learns generative statistical models of cell spatial organization from fluorescence microscopy images. A trained model captures not just a single representative cell shape, but the full distribution of variation — in overall cell architecture, organelle count, organelle size and shape, and spatial positioning — across a population of imaged cells.
Devin Sullivan’s PhD thesis in the Murphy Lab at Carnegie Mellon University extends generative statistical modeling of subcellular organization into three dimensions and across time. Building on the CellOrganizer framework, the work learns models of cell geometry, organelle morphology, and spatial distribution directly from fluorescence microscopy images — and introduces temporal modeling to capture how cellular organization evolves dynamically.
The thesis addresses a core challenge in computational cell biology: how to represent the spatial organization of a cell in a form that is both statistically rigorous and practically useful for downstream simulation. Key contributions include:
