Single-cell analysis for prognosis and tumor biology assessment in lung adenocarcinoma

Spatial organization of the tumor microenvironment (TME) has prognostic relevance in lung adenocarcinoma (LUAD), but most spatial profiling methods are costly and not routinely available. We aimed to develop and validate an automated, interpretable AI-based pipeline for extracting single-cell and spatial TME parameters from routine hematoxylin and eosin (H&E)-stained whole-slide images of resected LUAD. Five independent cohorts of neoadjuvant therapy-naïve patients with resected LUAD were analyzed: TCGA, PLCO, NLST, KAM, and UKK, comprising 2791 whole-slide images from 1355 patients. The pipeline performed tissue detection, artifact quality control, tumor and stroma segmentation, and single-cell detection/classification. Cell densities, relative proportions, stromal cellularity, tumor cell–lymphocyte distance parameters, k-means-derived patient clusters, and cellular-neighborhood meta-clusters were extracted. Prognostic associations were assessed using Cox regression and Kaplan–Meier analyses across available survival endpoints, including overall, cancer-specific, and progression-free survival. High lymphocyte densities in tumor and tumor stroma were consistently associated with favorable outcomes, whereas increased stromal granulocyte, connective tissue cell, and plasma cell proportions were associated with adverse prognosis. Tumor cells lacking nearby lymphocytes, particularly within short distances, showed strong adverse prognostic value, highlighting the importance of spatial immune-tumor interactions. Integrative k-means clustering identified biologically coherent risk groups, including lymphocyte-rich favorable and granulocyte-rich adverse phenotypes, with validation across cohorts. Cellular-neighborhood analysis revealed ten spatial meta-clusters representing tumor core, stroma, and tumor–stroma interface patterns; several meta-clusters showed independent prognostic value. An automated AI pipeline can extract interpretable, prognostically relevant single-cell and spatial TME features from routine H&E slides in resected LUAD. These findings suggest that clinically meaningful spatial immune and stromal profiling may be achievable at scale without multiplexed imaging, supporting future prospective validation for routine risk stratification.