SHIELD: A weakly supervised graph attention neural network for decoding disease-relevant cell-cell interactions

Multiplexed tissue imaging enables detailed study of cell-cell interactions in disease, yet systematic, interpretable, and supervised computational methods for inferring such interactions remain scarce. We present SHIELD (spatially enhanced immune landscape decoding), a graph attention network framework that quantifies disease-relevant cell-cell interactions through learned attention scores, without relying on prior biological assumptions such as ligand-receptor databases. Validated across three multiplexed tissue imaging datasets-hepatocellular carcinoma (HCC), colorectal cancer (CRC), and type 1 diabetes (T1D)-SHIELD identifies rare mucosal-associated invariant T (MAIT) cell-macrophage interactions in HCC, suppressive CD8⁺ T cell-macrophage interactions enriched in CRC non-responders, and β cell interactions with cytotoxic and helper T cells across T1D disease stages. In all contexts, SHIELD reconstructs known and biologically meaningful interactions, offering a robust, interpretable tool for spatial tissue analysis and data-driven mechanistic discovery.