Choragraph: A deep-learning approach for the analysis of spatial proteomics reveals subcellular Arabidopsis protein trafficking routes and multi-residency

Abstract

Spatial subcellular proteomics provides key insights into subcellular organization but is frequently constrained by missing data values and an inability to robustly classify dual-localized proteins. To address these analytical bottlenecks, we introduce Choragraph, a deep-learning framework that uses an ensemble of deep neural networks incorporating whole-proteome cross-attention and Bayesian variational inference (BVI). Choragraph has two concurrent aims: it provides context-dependent reconstruction of missing proteomic values and predicts proteins subcellular compartment in a manner natively aware of multi-localisation. We applied Choragraph to a comprehensive Arabidopsis thaliana hyperLOPIT dataset containing 84 fractions across eight replicate LOPIT experiments. By successfully reconstructing profiles with up to 35% missing values, Choragraph incorporated over 2,000 low-abundance proteins that traditional methods would exclude. The models confidently classified 87% to 94% of singly-localised data-sufficient proteins across 14 subcellular compartments with a macro F1 score of 0.917, outperforming conventional classifiers. Crucially, Choragraph also identified over 1,000 dual-localized proteins, mapping continuous trafficking trails along the secretory pathway, highlighting functional zonation and membrane contact sites. To ensure accessibility, all data, predictions of subcellular localisation, and interactive 2D UMAP visualizations are available via an installation-free web application at choragraph.org. This framework provides a high-resolution, user-friendly resource that advances research capacity to explore subcellular location and dynamics