Graph neural network modeling of spatial tumor-immune interactions identifies prognostic cellular niches in non‑small cell lung cancer.

The spatial organization of immune and tumor cells within the tumor microenvironment (TME) has a critical influence on antitumor immunity and patient survival. However, hand-engineered metrics such as cell densities or pairwise proximity scores fail to capture the complexity of local cell-cell interactions. Understanding these higher-order spatial patterns and their relation to patient outcomes is especially important for non-small cell lung cancer (NSCLC), the deadliest cancer worldwide. Here, we elucidate the NSCLC TME using a graph neural network (GNN)-based framework to model spatially localized cellular neighborhoods in multiplex immunofluorescence data from a clinical cohort of 506 patients. The GNN predicted patient survival with high accuracy (concordance index: 0.82) and remained a significant prognostic factor when adjusted for clinical covariates. Interpretability analyses revealed that specific combinations of cell types, particularly involving CD8+ T cells, PD-L1+ immune cells, and FOXP3+ regulatory T cells, modulated predictions depending on their spatial context. In-silico manipulation experiments applied to the trained GNN, used here as an interpretable surrogate model, suggested that the impact of CD8+ cells on survival were estimated as favorable when in direct tumor contact and less favorable when adjacent to immunosuppressive cells. Latent-space clustering identified distinct TME states predictive of outcome, reflecting varying balances of immune activation and evasion. Our approach underscores the prognostic significance of spatially resolved immune-tumor interactions, providing a blueprint for developing next-generation spatial biomarkers to guide precision treatment strategies in NSCLC.