The paper outlines the creation of a global, multiscale map of human subcellular architecture in U2OS osteosarcoma cells, integrating two complementary data sources: biophysical protein interactions determined by affinity purification–mass spectrometry (AP–MS) and protein localization images from immunofluorescence (IF). The researchers employed a self-supervised machine learning model to fuse these modalities, resolving 275 distinct molecular assemblies spanning a broad range of physical sizes, which they systematically validated using whole-cell size-exclusion chromatography (SEC–MS). This reference map was annotated with the help of large language models (LLMs) and applied to diverse biological problems, including 3D structural modeling, assigning unexpected functions to hundreds of proteins, analyzing cell-type specificity and multi-localization, and interpreting paediatric cancer mutations that converge on specific protein assemblies. The study culminates in a reference platform and open-source toolkit for structural and functional cell biology.
References:
- Schaffer L V, Hu M, Qian G, et al. Multimodal cell maps as a foundation for structural and functional genomics[J]. Nature, 2025: 1-10.
