SR2P: an efficient stacking method to predict protein abundance from gene expression in spatial transcriptomics data

Abstract

Spatial transcriptomics data are largely available with RNA expression alone, limiting the detection of cell states defined by surface protein abundance. The lack of multi-omics spatial data limits the ability to identify immune cells and their signaling in the tumor microenvironment, as most solid tumors are immunologically poor and exhibit protein-RNA abundance discordance in critical immune cell surface markers. Although emerging technologies enable spatial multi-omics profiling, technical and cost constraints remain a hurdle. We introduce SR2P, a stacking-based machine-learning framework for predicting spatial protein abundance from RNA expression. SR2P integrates 11 complementary predictive models and consistently outperforms existing methods across multiple spatial multi-omics benchmark. We showcased an application of SR2P recovered macrophage-enriched regions and identified potential immune markers associated with therapeutic response from head-and-neck squamous cell carcinoma patients. SR2P enables protein-abundance inference from RNA-only spatial data, extending the analytical capabilities of current spatial platforms for studies of tumor immunology.