SPgen: Proteome-wide Spatial Proteomics generation using multi-modality foundation models
TL;DR
SPgen integrates protein sequences, functional annotations, transcriptomic profiles, and spatial information to learn transferable representations that enable inference beyond experimentally profiled proteins, and accurately reconstructs measured spatial patterns, reduces measurement noise, and enables proteome-wide spatial prediction.
Abstract
Spatial proteomics (SP) measures the spatial distribution of proteins within tissues, providing important insights into tissue function, disease, and therapeutic response. However, current SP technologies profile only a small fraction of the proteome and are limited by cost and measurement noise. Recent AI approaches enable predicting spatial protein expression from transcriptomic or histopathological data, but are typically restricted to paired datasets covering only tens of proteins, limiting their ability to generalize beyond experimentally measured protein panels. Here we present SPgen, a multi-modal foundation-model framework for proteome-wide spatial protein prediction. SPgen integrates protein sequences, functional annotations, transcriptomic profiles, and spatial information to learn transferable representations that enable inference beyond experimentally profiled proteins. Across diverse spatial proteomics datasets, SPgen accurately reconstructs measured spatial patterns, reduces measurement noise, and enables proteome-wide spatial prediction.