Identification of a Conserved Dedifferentiated Cancer Cell State Stratifying Poor Survival in Lung Adenocarcinoma

Cancer cell plasticity underlies therapeutic resistance, yet its determinants remain poorly understood. While genetic mutations drive carcinogenesis, the non-malignant cell-of-origin state also dictates tumor behavior. Using single-cell transcriptomics, we characterized cancer cell subpopulations in lung adenocarcinoma by integrating non-malignant alveolar type 2 and cancer cell transcriptional states. In clinical trial bulk transcriptomes, gene signatures of four malignant subpopulations (Groups 0–3) revealed that Group 1 is strongly associated with significantly worse overall survival and exhibits a highly aggressive phenotype with stemness, hypoxia, epithelial-mesenchymal transition and intermicrovillar adhesion complex dysregulation. An ElasticNet Cox model trained on pooled cohorts (IMpower150 + OAK, N = 1,572, events = 1,205) yielded a 35-gene risk signature. In locked-model validation across held-out cohorts (IMpower110, IMpower130, BIRCH, and POPLAR), high Group 1 risk scores demonstrated strong prognostic value for elevated mortality (pooled random-effects hazard ratio [HR] = 1.326 per 1 SD increase, 95% CI: 1.218 to 1.443; p < 0.001), and tertile contrast confirmed pronounced survival separation (HR = 1.971, 95% CI: 1.486 to 2.614). In multivariable Cox regression, the signature was independent of age, sex, treatment regimen, smoking status and geographic location. High-risk patients across immunotherapy and chemotherapy exhibited strong activation of pro-survival and oncogenic networks (e.g., EGFR, PI3K and MAPK), whereas low-risk patients showed higher immune-related pathway activity (e.g., JAK-STAT). These findings suggest that cancer cells co-opt conserved developmental programs associated with poor outcomes regardless of treatment. The 35-gene Group 1 signature is a robust prognostic biomarker across immunotherapy and chemotherapy and provides a foundation for rational combination strategies.