Contrastive learning uncovers cellular interactions and morphologies in the tumor microenvironment of lung adenocarcinoma linked to immunotherapy response.

Lung adenocarcinoma (LUAD) presents diverse histomorphological features within the tumor microenvironment (TME) that influence prognosis and response to immunotherapy. Leveraging contrastive learning, we developed an unbiased atlas of cell neighborhoods to systematically explore the LUAD microenvironment at the cellular scale and investigate how these neighborhoods combine to form histologic patterns. This multiscale approach enables a comprehensive understanding of both cell-specific interactions and broader histologic architecture in LUAD. Our analysis identified distinct histomorphological phenotype clusters of cellular neighborhoods (cn-HPCs) with prognostic significance. For example, cn-HPC 0 was associated with immune activation and favorable survival, while cn-HPC 23 was enriched in necrotic, immune-excluded regions and aligned with poorer outcomes. Through associations with immunophenotypes, co-expressed gene modules, and pathway enrichment, we found that cn-HPCs reflect underlying processes of immune modulation, cellular growth, and inflammation, offering insights into the functional landscape of LUAD. In an independent LUAD immunotherapy cohort, we found that cn-HPC 23 showed promising predictive value for treatment response. These findings suggest that fine-grained spatial profiling of the TME may offer complementary value for identifying patients more likely to benefit from immunotherapy. Taken together, our study highlights the potential of cn-HPCs to bridge histopathological patterns and clinical outcomes, supporting their relevance for prognosis and treatment stratification in LUAD.