Multi-cohort machine learning identifies a ferroptosis-linked prognostic signature in lung adenocarcinoma

Introduction Lung adenocarcinoma (LUAD) shows marked outcome heterogeneity within clinicopathological stage groups. This study developed and externally validated a ferroptosis-linked transcriptomic risk model using a leakage-controlled multi- cohort survival-learning framework and characterized its immune context. Methods Candidate genes were defined by combining FerrDb V3-annotated genes within a ferroptosis-associated weighted gene co-expression network analysis (WGCNA) module with a filtered WGCNA discovery branch derived from the independent GSE81089 cohort. Model development used TCGA-LUAD, GSE31210, GSE72094, and GSE136961 (1,149 patients; 344 overall-survival events) with bagged cross-cohort Cox screening, leave-one-cohort-out (LOCO) stability locking, and benchmarking of 150 survival-learning configurations. External evaluation used the GSE50081, GSE68465, and GSE30219 cohorts (912 patients; 507 events). Results The development-selected extra survival trees configuration reached a mean leave-one-cohort-out Uno’s C-index of 0.723. The highest cohort-specific C-indices among the prespecified candidate configurations were 0.611, 0.691, and 0.699 and arose from different model configurations. Applying the same development-selected configuration to all three external cohorts yielded Harrell’s C-indices of 0.605, 0.675, and 0.682 and 5-year Uno’s C-indices of 0.605, 0.678, and 0.696. In TCGA-LUAD, the stored out-of-fold molecular score remained prognostic after TNM stage adjustment (hazard ratio per standard deviation 2.32, 95% confidence interval: 1.39−3.87; p = 0.0013). Discussion At 3 years and 5 years, the combined TNM-plus-score models showed close calibration and modest gains in discrimination, while decision-curve analysis identified positive incremental net benefit only over restricted threshold ranges. Low-risk tumors were enriched for interferon, complement, and immune-cell programs. These computational findings require further prospective assay-level and experimental validation before clinical implementation.