Proximity-Informed Graph Learning Defines Spatial Protein Communities for Tumor-Associated Proximity Antigen Discovery

Abstract The spatial organization of membrane proteins is an underexplored dimension of cell-surface biology. Spatial proximity shapes cellular function and therapeutic targetability, yet efforts to identify tumor-associated antigens (TAAs) have largely focused on expression alone. Here, we developed an industrialized surface-protein proximity mapping workflow to interrogate TAAs within their membrane microenvironments. In the process, we generated 248 proximity maps across 12 receptor tyrosine kinases (RTKs) and 28 tumor cell systems. This proximity atlas enabled two advances: first, MetaMap, a correlation-based analytical framework that defines spatial protein communities and infers non-targeted proximal proteins from reproducible proximity signatures; and second, tumor-associated proximity antigens (TAPAs), a conceptual class of co-targets defined by disease-specific spatial proximity to TAAs rather than expression alone. Applying these proximity-derived relationships within a multimodal prioritization framework, we identified and validated an EGFR×CDCP1 TAA-TAPA pair that enhanced tumor cell killing across therapeutic modalities. By integrating spatial organization with multimodal data, this work expands the design space for precision-guided therapeutic strategies.