Title: Towards a systematic understanding of human cells through data-driven models of spatial proteomics

Abstract: Cells are spatially organized multi-scaled systems where precise protein localization underpins cellular identity, function, and dynamic response to perturbations. Spatial proteomics provides a powerful framework to systematically characterize this organization and elucidate underlying biological mechanisms at subcellular resolution.

This thesis presents a body of work that advances the study of spatial proteomics through computational modeling, large-scale imaging, and open-source tool development. It begins with the design and analysis of a citizen science competition for predicting protein subcellular localization from confocal microscopy images, establishing one of the first frameworks capable of capturing single-cell heterogeneity and subcellular dynamics across 19 compartments in 20+ cell lines. A subsequent chapter introduces a shape-informed spatial coordinate system that integrates signal processing and machine learning to map protein localization across variable morphologies in more than 1M single cells, revealing that shape-dependent spatial variability can indicate distinct cellular states beyond cell cycle effects. This serves as the first systematic map of cell and nuclear shape with molecular information at the organelle, pathway and protein levels. The 3rd work then explores the spatial proteomic landscape of SARS-CoV-2 infection through high-throughput immunofluorescence imaging, identifying over 100 proteins with altered localization or abundance and nominating drug repurposing candidates through phenotypic screening. Finally, two methodological contributions focus on enabling broader accessibility to bioimage modeling: an interactive web-based framework for real-time model training and annotation, and a generative model for virtual cell painting, enabling label-free organelle prediction from transmitted light images. Collectively, these works provide foundational knowledge toward a multimodal, multiscale virtual cell model, exemplified by a data-driven, hierarchical representation that integrates protein localization and interaction to investigate emergent subcellular assemblies and uncover previously uncharacterized cellular functions.

Please contact Madelyn Bernstein for the Zoom link