Morphology-robust quantification of subcellular organization in complex cells

Quantitative analysis of subcellular protein organization is often confounded by variation in cell morphology, limiting the identification and interpretation of localization patterns in fluorescence microscopy data from morphologically complex cells, such as neurons and glia. We introduce CellAligner, an unsupervised framework that uses fused unbalanced Gromov-Wasserstein couplings to map protein distributions from morphologically distinct cells into shared anchor-cell geometries, enabling morphology-robust comparison of subcellular localization. In neuronal imaging benchmarks, applying current image-analysis methods (CellProfiler, Cytoself, Paired Cell Inpainting) to CellAligner's anchor-cell representations substantially reduced morphology-associated confounding while approximately doubling their multiclass MCC for localization classification. We demonstrate its biological utility by identifying U18666A-induced lysosomal trafficking defects in human iPSC-derived neurons. To scale the approach, we developed dCellAligner-OT, a fast deep metric learning model that approximates CellAligner's optimal transport distances and anchor-cell representations, enabling atlas-scale analyses. CellAligner provides a general framework for morphology-robust analysis of subcellular organization in complex cellular systems.